Content from Introduction
Last updated on 2025-10-15 | Edit this page
Overview
Questions
- Question 1
- Question 2
Objectives
- Explain the need for image processing and analysis in the context of biological research.
- Explain the concept of a workflow in image processing and analysis and list common building blocks.
As computer systems have become faster and more powerful, and cameras and other imaging systems have become commonplace in many other areas of life, the need has grown for researchers to be able to process and analyse image data. Considering the large volumes of data that can be involved - high-resolution images that take up a lot of disk space/virtual memory, and/or collections of many images that must be processed together - and the time-consuming and error-prone nature of manual processing, it can be advantageous or even necessary for this processing and analysis to be automated as a computer program.
This lesson introduces an open source toolkit for processing image
data: the Python programming language and the scikit-image
(skimage) library. With careful experimental design,
Python code can be a powerful instrument in answering many different
kinds of questions.
Uses of Image Processing in Research
Automated processing can be used to analyse many different properties of an image, including the distribution and change in colours in the image, the number, size, position, orientation, and shape of objects in the image, and even - when combined with machine learning techniques for object recognition - the type of objects in the image.
Some examples of image processing methods applied in research include:
- Imaging a black hole
- Segmentation of liver and vessels from CT images
- Monitoring wading birds in the Everglades using drones (Blog article summarizing the paper)
- Estimating the population of emperor penguins
- Global-scale analysis of marine plankton diversity
With this lesson, we aim to provide a thorough grounding in the fundamental concepts and skills of working with image data in Python. Most of the examples used in this lesson focus on one particular class of image processing technique, morphometrics, but what you will learn can be used to solve a much wider range of problems.
Morphometrics
Morphometrics involves counting the number of objects in an image, analyzing the size of the objects, or analyzing the shape of the objects. For example, we might be interested in automatically counting the number of bacterial colonies growing in a Petri dish, as shown in this image:
We could use image processing to find the colonies, count them, and then highlight their locations on the original image, resulting in an image like this:
Why write a program to do that?
Note that you can easily manually count the number of bacteria colonies shown in the morphometric example above. Why should we learn how to write a Python program to do a task we could easily perform with our own eyes? There are at least two reasons to learn how to perform tasks like these with Python and scikit-image:
- What if there are many more bacteria colonies in the Petri dish? For example, suppose the image looked like this:
Manually counting the colonies in that image would present more of a challenge. A Python program using scikit-image could count the number of colonies more accurately, and much more quickly, than a human could.
- What if you have hundreds, or thousands, of images to consider? Imagine having to manually count colonies on several thousand images like those above. A Python program using scikit-image could move through all of the images in seconds; how long would a graduate student require to do the task? Which process would be more accurate and repeatable?
As you can see, the simple image processing / computer vision techniques you will learn during this workshop can be very valuable tools for scientific research.
As we move through this workshop, we will learn image analysis methods useful for many different scientific problems. These will be linked together and applied to a real problem in the final end-of-workshop capstone challenge.
Let’s get started, by learning some basics about how images are represented and stored digitally.
- Simple Python and scikit-image techniques can be used to solve genuine image analysis problems.
- Morphometric problems involve the number, shape, and / or size of the objects in an image.
Content from Image Basics
Last updated on 2025-10-15 | Edit this page
Overview
Questions
- Question 1
Objectives
Describe the main differences between typical fluorescence bioimages and scientific (like histological staining) and non-scientific (like camera pictures) RGB images.
Explain how some aspects of image origin and formation can influence downstream analysis.
Load images into Python represented as n-dimensional arrays via BioIO
Display pixel values from a NumPy array
Extract and interpret an image’s shape
Find an image’s min, max, and mean pixel value
Explain the significance of bit depth in images and how it affects image quality and memory footprint
-
Explain how bit depth and data type (dtype) determine the range of valid pixel values in an image.
- Provide examples of valid values for different NumPy array data type.
- Explain the link between bit depth, data type and possible values present in an image
- Present examples of bit depth, data type, shape and values for common bioimages.
The images we see on hard copy, view with our electronic devices, or process with our programs are represented and stored in the computer as numeric abstractions, approximations of what we see with our eyes in the real world. Before we begin to learn how to process images with Python programs, we need to spend some time understanding how these abstractions work.
Feel free to make use of the available cheat-sheet as a guide for the rest of the course material. View it online, share it, or print the PDF!
Pixels
It is important to realise that images are stored as rectangular arrays of hundreds, thousands, or millions of discrete “picture elements,” otherwise known as pixels. Each pixel can be thought of as a single square point of coloured light.
For example, consider this image of a maize seedling, with a square area designated by a red box:
Now, if we zoomed in close enough to see the pixels in the red box, we would see something like this:
Note that each square in the enlarged image area - each pixel - is all one colour, but that each pixel can have a different colour from its neighbors. Viewed from a distance, these pixels seem to blend together to form the image we see.
Real-world images are typically made up of a vast number of pixels, and each of these pixels is one of potentially millions of colours. While we will deal with pictures of such complexity in this lesson, let’s start our exploration with just 15 pixels in a 5 x 3 matrix with 2 colours, and work our way up to that complexity.
Matrices, arrays, images and pixels
A matrix is a mathematical concept - numbers evenly arranged in a rectangle. This can be a two-dimensional rectangle, like the shape of the screen you’re looking at now. Or it could be a three-dimensional equivalent, a cuboid, or have even more dimensions, but always keeping the evenly spaced arrangement of numbers. In computing, an array refers to a structure in the computer’s memory where data is stored in evenly spaced elements. This is strongly analogous to a matrix. A NumPy array is a type of variable (a simpler example of a type is an integer). For our purposes, the distinction between matrices and arrays is not important, we don’t really care how the computer arranges our data in its memory. The important thing is that the computer stores values describing the pixels in images, as arrays. And the terms matrix and array will be used interchangeably.
Loading images
As noted, images we want to analyze (process) with Python are loaded into arrays. There are multiple ways to load images. In this lesson, we use imageio, a Python library for reading (loading) and writing (saving) image data, and more specifically its version 3. But, really, we could use any image loader which would return a NumPy array.
The v3 module of imageio (imageio.v3) is
imported as iio (see note in the next section). Version 3
of imageio has the benefit of supporting nD (multidimensional) image
data natively (think of volumes, movies).
Let us load our image data from disk using the imread
function from the imageio.v3 module.
OUTPUT
<class 'numpy.ndarray'>Note that, using the same image loader or a different one, we could also read in remotely hosted data.
Why not use
skimage.io.imread()?
The scikit-image library has its own function to read an image, so
you might be asking why we don’t use it here. Actually,
skimage.io.imread() uses iio.imread()
internally when loading an image into Python. It is certainly something
you may use as you see fit in your own code. In this lesson, we use the
imageio library to read or write images, while scikit-image is dedicated
to performing operations on the images. Using imageio gives us more
flexibility, especially when it comes to handling metadata.
Beyond NumPy arrays
Beyond NumPy arrays, there exist other types of variables which are
array-like. Notably, pandas.DataFrame
and xarray.DataArray
can hold labeled, tabular data. These are not natively supported in
scikit-image, the scientific toolkit we use in this lesson for
processing image data. However, data stored in these types can be
converted to numpy.ndarray with certain assumptions (see
pandas.DataFrame.to_numpy() and
xarray.DataArray.data). Particularly, these conversions
ignore the sampling coordinates (DataFrame.index,
DataFrame.columns, or DataArray.coords), which
may result in misrepresented data, for instance, when the original data
points are irregularly spaced.
Working with pixels
First, let us add the necessary imports:
PYTHON
"""Python libraries for learning and performing image processing."""
import ipympl
import matplotlib.pyplot as plt
import numpy as np
import skimage as skiImport statements in Python
In Python, the import statement is used to load
additional functionality into a program. This is necessary when we want
our code to do something more specialised, which cannot easily be
achieved with the limited set of basic tools and data structures
available in the default Python environment.
Additional functionality can be loaded as a single function or
object, a module defining several of these, or a library containing many
modules. You will encounter several different forms of
import statement.
PYTHON
import skimage # form 1, load whole skimage library
import skimage.draw # form 2, load skimage.draw module only
from skimage.draw import disk # form 3, load only the disk function
import skimage as ski # form 4, load all of skimage into an object called skiIn the example above, form 1 loads the entire scikit-image library
into the program as an object. Individual modules of the library are
then available within that object, e.g., to access the disk
function used in the drawing episode, you
would write skimage.draw.disk().
Form 2 loads only the draw module of
skimage into the program. The syntax needed to use the
module remains unchanged: to access the disk function, we
would use the same function call as given for form 1.
Form 3 can be used to import only a specific function/class from a
library/module. Unlike the other forms, when this approach is used, the
imported function or class can be called by its name only, without
prefixing it with the name of the library/module from which it was
loaded, i.e., disk() instead of
skimage.draw.disk() using the example above. One hazard of
this form is that importing like this will overwrite any object with the
same name that was defined/imported earlier in the program, i.e., the
example above would replace any existing object called disk
with the disk function from skimage.draw.
Finally, the as keyword can be used when importing, to
define a name to be used as shorthand for the library/module being
imported. This name is referred to as an alias. Typically, using an
alias (such as np for the NumPy library) saves us a little
typing. You may see as combined with any of the other first
three forms of import statements.
Which form is used often depends on the size and number of additional tools being loaded into the program.
Now that we have our libraries loaded, we will run a Jupyter Magic Command that will ensure our images display in our Jupyter document with pixel information that will help us more efficiently run commands later in the session.
With that taken care of, let us display the image we have loaded,
using the imshow function from the
matplotlib.pyplot module.
You might be thinking, “That does look vaguely like an eight, and I see two colours but how can that be only 15 pixels”. The display of the eight you see does use a lot more screen pixels to display our eight so large, but that does not mean there is information for all those screen pixels in the file. All those extra pixels are a consequence of our viewer creating additional pixels through interpolation. It could have just displayed it as a tiny image using only 15 screen pixels if the viewer was designed differently.
While many image file formats contain descriptive metadata that can
be essential, the bulk of a picture file is just arrays of numeric
information that, when interpreted according to a certain rule set,
become recognizable as an image to us. Our image of an eight is no
exception, and imageio.v3 stored that image data in an
array of arrays making a 5 x 3 matrix of 15 pixels. We can demonstrate
that by calling on the shape property of our image variable and see the
matrix by printing our image variable to the screen.
OUTPUT
(5, 3)
[[0. 0. 0.]
[0. 1. 0.]
[0. 0. 0.]
[0. 1. 0.]
[0. 0. 0.]]Thus if we have tools that will allow us to manipulate these arrays
of numbers, we can manipulate the image. The NumPy library can be
particularly useful here, so let’s try that out using NumPy array
slicing. Notice that the default behavior of the imshow
function appended row and column numbers that will be helpful to us as
we try to address individual or groups of pixels. First let’s load
another copy of our eight, and then make it look like a zero.
To make it look like a zero, we need to change the number underlying the centremost pixel to be 1. With the help of those row and column headers, at this small scale we can determine the centre pixel is in row labeled 2 and column labeled 1. Using array slicing, we can then address and assign a new value to that position.
PYTHON
zero = iio.imread(uri="data/eight.tif")
zero[2, 1]= 1.0
# The following line of code creates a new figure for imshow to use in displaying our output.
fig, ax = plt.subplots()
ax.imshow(zero)
print(zero)OUTPUT
[[0. 0. 0.]
[0. 1. 0.]
[0. 1. 0.]
[0. 1. 0.]
[0. 0. 0.]]
Coordinate system
When we process images, we can access, examine, and / or change the colour of any pixel we wish. To do this, we need some convention on how to access pixels individually; a way to give each one a name, or an address of a sort.
The most common manner to do this, and the one we will use in our programs, is to assign a modified Cartesian coordinate system to the image. The coordinate system we usually see in mathematics has a horizontal x-axis and a vertical y-axis, like this:
The modified coordinate system used for our images will have only positive coordinates, the origin will be in the upper left corner instead of the centre, and y coordinate values will get larger as they go down instead of up, like this:
This is called a left-hand coordinate system. If you hold your left hand in front of your face and point your thumb at the floor, your extended index finger will correspond to the x-axis while your thumb represents the y-axis.
Until you have worked with images for a while, the most common mistake that you will make with coordinates is to forget that y coordinates get larger as they go down instead of up as in a normal Cartesian coordinate system. Consequently, it may be helpful to think in terms of counting down rows (r) for the y-axis and across columns (c) for the x-axis. This can be especially helpful in cases where you need to transpose image viewer data provided in x,y format to y,x format. Thus, we will use cx and ry where appropriate to help bridge these two approaches.
Changing Pixel Values (5 min)
Load another copy of eight named five, and then change the value of pixels so you have what looks like a 5 instead of an 8. Display the image and print out the matrix as well.
More colours
Up to now, we only had a 2 colour matrix, but we can have more if we use other numbers or fractions. One common way is to use the numbers between 0 and 255 to allow for 256 different colours or 256 different levels of grey. Let’s try that out.
PYTHON
# make a copy of eight
three_colours = iio.imread(uri="data/eight.tif")
# multiply the whole matrix by 128
three_colours = three_colours * 128
# set the middle row (index 2) to the value of 255.,
# so you end up with the values 0., 128., and 255.
three_colours[2, :] = 255.
fig, ax = plt.subplots()
ax.imshow(three_colours)
print(three_colours)
We now have 3 colours, but are they the three colours you expected? They all appear to be on a continuum of dark purple on the low end and yellow on the high end. This is a consequence of the default colour map (cmap) in this library. You can think of a colour map as an association or mapping of numbers to a specific colour. However, the goal here is not to have one number for every possible colour, but rather to have a continuum of colours that demonstrate relative intensity. In our specific case here for example, 255 or the highest intensity is mapped to yellow, and 0 or the lowest intensity is mapped to a dark purple. The best colour map for your data will vary and there are many options built in, but this default selection was not arbitrary. A lot of science went into making this the default due to its robustness when it comes to how the human mind interprets relative colour values, grey-scale printability, and colour-blind friendliness (You can read more about this default colour map in a Matplotlib tutorial and an explanatory article by the authors). Thus it is a good place to start, and you should change it only with purpose and forethought. For now, let’s see how you can do that using an alternative map you have likely seen before where it will be even easier to see it as a mapped continuum of intensities: greyscale.
Above we have exactly the same underlying data matrix, but in greyscale. Zero maps to black, 255 maps to white, and 128 maps to medium grey. Here we only have a single channel in the data and utilize a grayscale color map to represent the luminance, or intensity of the data and correspondingly this channel is referred to as the luminance channel.
Even more colours
This is all well and good at this scale, but what happens when we instead have a picture of a natural landscape that contains millions of colours. Having a one to one mapping of number to colour like this would be inefficient and make adjustments and building tools to do so very difficult. Rather than larger numbers, the solution is to have more numbers in more dimensions. Storing the numbers in a multi-dimensional matrix where each colour or property like transparency is associated with its own dimension allows for individual contributions to a pixel to be adjusted independently. This ability to manipulate properties of groups of pixels separately will be key to certain techniques explored in later chapters of this lesson. To get started let’s see an example of how different dimensions of information combine to produce a set of pixels using a 4 x 4 matrix with 3 dimensions for the colours red, green, and blue. Rather than loading it from a file, we will generate this example using NumPy.
PYTHON
# set the random seed so we all get the same matrix
pseudorandomizer = np.random.RandomState(2021)
# create a 4 × 4 checkerboard of random colours
checkerboard = pseudorandomizer.randint(0, 255, size=(4, 4, 3))
# restore the default map as you show the image
fig, ax = plt.subplots()
ax.imshow(checkerboard)
# display the arrays
print(checkerboard)OUTPUT
[[[116 85 57]
[128 109 94]
[214 44 62]
[219 157 21]]
[[ 93 152 140]
[246 198 102]
[ 70 33 101]
[ 7 1 110]]
[[225 124 229]
[154 194 176]
[227 63 49]
[144 178 54]]
[[123 180 93]
[120 5 49]
[166 234 142]
[ 71 85 70]]]
Previously we had one number being mapped to one colour or intensity. Now we are combining the effect of 3 numbers to arrive at a single colour value. Let’s see an example of that using the blue square at the end of the second row, which has the index [1, 3].
PYTHON
# extract all the colour information for the blue square
upper_right_square = checkerboard[1, 3, :]
upper_right_squareThis outputs: array([ 7, 1, 110]) The integers in order represent Red, Green, and Blue. Looking at the 3 values and knowing how they map, can help us understand why it is blue. If we divide each value by 255, which is the maximum, we can determine how much it is contributing relative to its maximum potential. Effectively, the red is at 7/255 or 2.8 percent of its potential, the green is at 1/255 or 0.4 percent, and blue is 110/255 or 43.1 percent of its potential. So when you mix those three intensities of colour, blue is winning by a wide margin, but the red and green still contribute to make it a slightly different shade of blue than 0,0,110 would be on its own.
These colours mapped to dimensions of the matrix may be referred to as channels. It may be helpful to display each of these channels independently, to help us understand what is happening. We can do that by multiplying our image array representation with a 1d matrix that has a one for the channel we want to keep and zeros for the rest.
If we look at the upper [1, 3] square in all three figures, we can see each of those colour contributions in action. Notice that there are several squares in the blue figure that look even more intensely blue than square [1, 3]. When all three channels are combined though, the blue light of those squares is being diluted by the relative strength of red and green being mixed in with them.
24-bit RGB colour
This last colour model we used, known as the RGB (Red, Green, Blue) model, is the most common.
As we saw, the RGB model is an additive colour model, which means that the primary colours are mixed together to form other colours. Most frequently, the amount of the primary colour added is represented as an integer in the closed range [0, 255] as seen in the example. Therefore, there are 256 discrete amounts of each primary colour that can be added to produce another colour. The number of discrete amounts of each colour, 256, corresponds to the number of bits used to hold the colour channel value, which is eight (28=256). Since we have three channels with 8 bits for each (8+8+8=24), this is called 24-bit colour depth.
Any particular colour in the RGB model can be expressed by a triplet of integers in [0, 255], representing the red, green, and blue channels, respectively. A larger number in a channel means that more of that primary colour is present.
Thinking about RGB colours (5 min)
Suppose that we represent colours as triples (r, g, b), where each of r, g, and b is an integer in [0, 255]. What colours are represented by each of these triples? (Try to answer these questions without reading further.)
- (255, 0, 0)
- (0, 255, 0)
- (0, 0, 255)
- (255, 255, 255)
- (0, 0, 0)
- (128, 128, 128)
- (255, 0, 0) represents red, because the red channel is maximised, while the other two channels have the minimum values.
- (0, 255, 0) represents green.
- (0, 0, 255) represents blue.
- (255, 255, 255) is a little harder. When we mix the maximum value of all three colour channels, we see the colour white.
- (0, 0, 0) represents the absence of all colour, or black.
- (128, 128, 128) represents a medium shade of gray. Note that the 24-bit RGB colour model provides at least 254 shades of gray, rather than only fifty.
Note that the RGB colour model may run contrary to your experience, especially if you have mixed primary colours of paint to create new colours. In the RGB model, the lack of any colour is black, while the maximum amount of each of the primary colours is white. With physical paint, we might start with a white base, and then add differing amounts of other paints to produce a darker shade.
After completing the previous challenge, we can look at some further examples of 24-bit RGB colours, in a visual way. The image in the next challenge shows some colour names, their 24-bit RGB triplet values, and the colour itself.
RGB colour table (optional, not included in timing)
We cannot really provide a complete table. To see why, answer this question: How many possible colours can be represented with the 24-bit RGB model?
There are 24 total bits in an RGB colour of this type, and each bit can be on or off, and so there are 224 = 16,777,216 possible colours with our additive, 24-bit RGB colour model.
Although 24-bit colour depth is common, there are other options. For example, we might have 8-bit colour (3 bits for red and green, but only 2 for blue, providing 8 × 8 × 4 = 256 colours) or 16-bit colour (4 bits for red, green, and blue, plus 4 more for transparency, providing 16 × 16 × 16 = 4096 colours, with 16 transparency levels each). There are colour depths with more than eight bits per channel, but as the human eye can only discern approximately 10 million different colours, these are not often used.
If you are using an older or inexpensive laptop screen or LCD monitor to view images, it may only support 18-bit colour, capable of displaying 64 × 64 × 64 = 262,144 colours. 24-bit colour images will be converted in some manner to 18-bit, and thus the colour quality you see will not match what is actually in the image.
We can combine our coordinate system with the 24-bit RGB colour model to gain a conceptual understanding of the images we will be working with. An image is a rectangular array of pixels, each with its own coordinate. Each pixel in the image is a square point of coloured light, where the colour is specified by a 24-bit RGB triplet. Such an image is an example of raster graphics.
Image formats
Although the images we will manipulate in our programs are conceptualised as rectangular arrays of RGB triplets, they are not necessarily created, stored, or transmitted in that format. There are several image formats we might encounter, and we should know the basics of at least of few of them. Some formats we might encounter, and their file extensions, are shown in this table:
| Format | Extension |
|---|---|
| Device-Independent Bitmap (BMP) | .bmp |
| Joint Photographic Experts Group (JPEG) | .jpg or .jpeg |
| Tagged Image File Format (TIFF) | .tif or .tiff |
BMP
The file format that comes closest to our preceding conceptualisation of images is the Device-Independent Bitmap, or BMP, file format. BMP files store raster graphics images as long sequences of binary-encoded numbers that specify the colour of each pixel in the image. Since computer files are one-dimensional structures, the pixel colours are stored one row at a time. That is, the first row of pixels (those with y-coordinate 0) are stored first, followed by the second row (those with y-coordinate 1), and so on. Depending on how it was created, a BMP image might have 8-bit, 16-bit, or 24-bit colour depth.
24-bit BMP images have a relatively simple file format, can be viewed and loaded across a wide variety of operating systems, and have high quality. However, BMP images are not compressed, resulting in very large file sizes for any useful image resolutions.
The idea of image compression is important to us for two reasons: first, compressed images have smaller file sizes, and are therefore easier to store and transmit; and second, compressed images may not have as much detail as their uncompressed counterparts, and so our programs may not be able to detect some important aspect if we are working with compressed images. Since compression is important to us, we should take a brief detour and discuss the concept.
Image compression
Before discussing additional formats, familiarity with image compression will be helpful. Let’s delve into that subject with a challenge. For this challenge, you will need to know about bits / bytes and how those are used to express computer storage capacities. If you already know, you can skip to the challenge below.
Bits and bytes
Before we talk specifically about images, we first need to understand how numbers are stored in a modern digital computer. When we think of a number, we do so using a decimal, or base-10 place-value number system. For example, a number like 659 is 6 × 102 + 5 × 101 + 9 × 100. Each digit in the number is multiplied by a power of 10, based on where it occurs, and there are 10 digits that can occur in each position (0, 1, 2, 3, 4, 5, 6, 7, 8, 9).
In principle, computers could be constructed to represent numbers in exactly the same way. But, the electronic circuits inside a computer are much easier to construct if we restrict the numeric base to only two, instead of 10. (It is easier for circuitry to tell the difference between two voltage levels than it is to differentiate among 10 levels.) So, values in a computer are stored using a binary, or base-2 place-value number system.
In this system, each symbol in a number is called a bit instead of a digit, and there are only two values for each bit (0 and 1). We might imagine a four-bit binary number, 1101. Using the same kind of place-value expansion as we did above for 659, we see that 1101 = 1 × 23 + 1 × 22 + 0 × 21 + 1 × 20, which if we do the math is 8 + 4 + 0 + 1, or 13 in decimal.
Internally, computers have a minimum number of bits that they work with at a given time: eight. A group of eight bits is called a byte. The amount of memory (RAM) and drive space our computers have is quantified by terms like Megabytes (MB), Gigabytes (GB), and Terabytes (TB). The following table provides more formal definitions for these terms.
| Unit | Abbreviation | Size |
|---|---|---|
| Kilobyte | KB | 1024 bytes |
| Megabyte | MB | 1024 KB |
| Gigabyte | GB | 1024 MB |
| Terabyte | TB | 1024 GB |
BMP image size (optional, not included in timing)
Imagine that we have a fairly large, but very boring image: a 5,000 × 5,000 pixel image composed of nothing but white pixels. If we used an uncompressed image format such as BMP, with the 24-bit RGB colour model, how much storage would be required for the file?
In such an image, there are 5,000 × 5,000 = 25,000,000 pixels, and 24 bits for each pixel, leading to 25,000,000 × 24 = 600,000,000 bits, or 75,000,000 bytes (71.5MB). That is quite a lot of space for a very uninteresting image!
Since image files can be very large, various compression schemes exist for saving (approximately) the same information while using less space. These compression techniques can be categorised as lossless or lossy.
Lossless compression
In lossless image compression, we apply some algorithm (i.e., a computerised procedure) to the image, resulting in a file that is significantly smaller than the uncompressed BMP file equivalent would be. Then, when we wish to load and view or process the image, our program reads the compressed file, and reverses the compression process, resulting in an image that is identical to the original. Nothing is lost in the process – hence the term “lossless.”
The general idea of lossless compression is to somehow detect long patterns of bytes in a file that are repeated over and over, and then assign a smaller bit pattern to represent the longer sample. Then, the compressed file is made up of the smaller patterns, rather than the larger ones, thus reducing the number of bytes required to save the file. The compressed file also contains a table of the substituted patterns and the originals, so when the file is decompressed it can be made identical to the original before compression.
To provide you with a concrete example, consider the 71.5 MB white BMP image discussed above. When put through the zip compression utility on Microsoft Windows, the resulting .zip file is only 72 KB in size! That is, the .zip version of the image is three orders of magnitude smaller than the original, and it can be decompressed into a file that is byte-for-byte the same as the original. Since the original is so repetitious - simply the same colour triplet repeated 25,000,000 times - the compression algorithm can dramatically reduce the size of the file.
If you work with .zip or .gz archives, you are dealing with lossless compression.
Lossy compression
Lossy compression takes the original image and discards some of the detail in it, resulting in a smaller file format. The goal is to only throw away detail that someone viewing the image would not notice. Many lossy compression schemes have adjustable levels of compression, so that the image creator can choose the amount of detail that is lost. The more detail that is sacrificed, the smaller the image files will be - but of course, the detail and richness of the image will be lower as well.
This is probably fine for images that are shown on Web pages or printed off on 4 × 6 photo paper, but may or may not be fine for scientific work. You will have to decide whether the loss of image quality and detail are important to your work, versus the space savings afforded by a lossy compression format.
It is important to understand that once an image is saved in a lossy compression format, the lost detail is just that - lost. I.e., unlike lossless formats, given an image saved in a lossy format, there is no way to reconstruct the original image in a byte-by-byte manner.
JPEG
JPEG images are perhaps the most commonly encountered digital images today. JPEG uses lossy compression, and the degree of compression can be tuned to your liking. It supports 24-bit colour depth, and since the format is so widely used, JPEG images can be viewed and manipulated easily on all computing platforms.
Examining actual image sizes (optional, not included in timing)
Let us see the effects of image compression on image size with actual images. The following script creates a square white image 5000 x 5000 pixels, and then saves it as a BMP and as a JPEG image.
PYTHON
dim = 5000
img = np.zeros((dim, dim, 3), dtype="uint8")
img.fill(255)
iio.imwrite(uri="data/ws.bmp", image=img)
iio.imwrite(uri="data/ws.jpg", image=img)Examine the file sizes of the two output files, ws.bmp
and ws.jpg. Does the BMP image size match our previous
prediction? How about the JPEG?
The BMP file, ws.bmp, is 75,000,054 bytes, which matches
our prediction very nicely. The JPEG file, ws.jpg, is
392,503 bytes, two orders of magnitude smaller than the bitmap
version.
Comparing lossless versus lossy compression (optional, not included in timing)
Let us see a hands-on example of lossless versus lossy compression.
Open a terminal (or Windows PowerShell) and navigate to the
data/ directory. The two output images, ws.bmp
and ws.jpg, should still be in the directory, along with
another image, tree.jpg.
We can apply lossless compression to any file by using the
zip command. Recall that the ws.bmp file
contains 75,000,054 bytes. Apply lossless compression to this image by
executing the following command: zip ws.zip ws.bmp
(Compress-Archive ws.bmp ws.zip with PowerShell). This
command tells the computer to create a new compressed file,
ws.zip, from the original bitmap image. Execute a similar
command on the tree JPEG file: zip tree.zip tree.jpg
(Compress-Archive tree.jpg tree.zip with PowerShell).
Having created the compressed file, use the ls -l
command (dir with PowerShell) to display the contents of
the directory. How big are the compressed files? How do those compare to
the size of ws.bmp and tree.jpg? What can you
conclude from the relative sizes?
Here is a partial directory listing, showing the sizes of the relevant files there:
OUTPUT
-rw-rw-r-- 1 diva diva 154344 Jun 18 08:32 tree.jpg
-rw-rw-r-- 1 diva diva 146049 Jun 18 08:53 tree.zip
-rw-rw-r-- 1 diva diva 75000054 Jun 18 08:51 ws.bmp
-rw-rw-r-- 1 diva diva 72986 Jun 18 08:53 ws.zipWe can see that the regularity of the bitmap image (remember, it is a
5,000 x 5,000 pixel image containing only white pixels) allows the
lossless compression scheme to compress the file quite effectively. On
the other hand, compressing tree.jpg does not create a much
smaller file; this is because the JPEG image was already in a compressed
format.
Here is an example showing how JPEG compression might impact image quality. Consider this image of several maize seedlings (scaled down here from 11,339 × 11,336 pixels in order to fit the display).
Now, let us zoom in and look at a small section of the label in the original, first in the uncompressed format:
Here is the same area of the image, but in JPEG format. We used a fairly aggressive compression parameter to make the JPEG, in order to illustrate the problems you might encounter with the format.
The JPEG image is of clearly inferior quality. It has less colour variation and noticeable pixelation. Quality differences become even more marked when one examines the colour histograms for each image. A histogram shows how often each colour value appears in an image. The histograms for the uncompressed (left) and compressed (right) images are shown below:
We learn how to make histograms such as these later on in the workshop. The differences in the colour histograms are even more apparent than in the images themselves; clearly the colours in the JPEG image are different from the uncompressed version.
If the quality settings for your JPEG images are high (and the compression rate therefore relatively low), the images may be of sufficient quality for your work. It all depends on how much quality you need, and what restrictions you have on image storage space. Another consideration may be where the images are stored. For example, if your images are stored in the cloud and therefore must be downloaded to your system before you use them, you may wish to use a compressed image format to speed up file transfer time.
PNG
PNG images are well suited for storing diagrams. It uses a lossless compression and is hence often used in web applications for non-photographic images. The format is able to store RGB and plain luminance (single channel, without an associated color) data, among others. Image data is stored row-wise and then, per row, a simple filter, like taking the difference of adjacent pixels, can be applied to increase the compressability of the data. The filtered data is then compressed in the next step and written out to the disk.
TIFF
TIFF images are popular with publishers, graphics designers, and photographers. TIFF images can be uncompressed, or compressed using either lossless or lossy compression schemes, depending on the settings used, and so TIFF images seem to have the benefits of both the BMP and JPEG formats. The main disadvantage of TIFF images (other than the size of images in the uncompressed version of the format) is that they are not universally readable by image viewing and manipulation software.
Metadata
JPEG and TIFF images support the inclusion of metadata in
images. Metadata is textual information that is contained within an
image file. Metadata holds information about the image itself, such as
when the image was captured, where it was captured, what type of camera
was used and with what settings, etc. We normally don’t see this
metadata when we view an image, but we can view it independently if we
wish to (see Accessing
Metadata, below). The important thing to be aware of at this
stage is that you cannot rely on the metadata of an image being fully
preserved when you use software to process that image. The image
reader/writer library that we use throughout this lesson,
imageio.v3, includes metadata when saving new images but
may fail to keep certain metadata fields. In any case, remember:
if metadata is important to you, take precautions to always
preserve the original files.
Accessing Metadata
imageio.v3 provides a way to display or explore the
metadata associated with an image. Metadata is served independently from
pixel data:
PYTHON
# read metadata
metadata = iio.immeta(uri="data/eight.tif")
# display the format-specific metadata
metadataOUTPUT
{'is_fluoview': False,
'is_nih': False,
'is_micromanager': False,
'is_ome': False,
'is_lsm': False,
'is_reduced': False,
'is_shaped': True,
'is_stk': False,
'is_tiled': False,
'is_mdgel': False,
'compression': <COMPRESSION.NONE: 1>,
'predictor': 1,
'is_mediacy': False,
'description': '{"shape": [5, 3]}',
'description1': '',
'is_imagej': False,
'software': 'tifffile.py',
'resolution_unit': 1,
'resolution': (1.0, 1.0, 'NONE')}Many popular image editing programs have built-in metadata viewing capabilities. A platform-independent open-source tool that allows users to read, write, and edit metadata is ExifTool. It can handle a wide range of file types and metadata formats but requires some technical knowledge to be used effectively. Other software exists that can help you handle metadata, e.g., Fiji and ImageMagick. You may want to explore these options if you need to work with the metadata of your images.
Summary of image formats used in this lesson
The following table summarises the characteristics of the BMP, JPEG, and TIFF image formats:
| Format | Compression | Metadata | Advantages | Disadvantages |
|---|---|---|---|---|
| BMP | None | None | Universally viewable, high quality | Large file sizes |
| JPEG | Lossy | Yes | Universally viewable, smaller file size | Detail may be lost |
| PNG | Lossless | Yes | Universally viewable, open standard, smaller file size | Metadata less flexible than TIFF, RGB only |
| TIFF | None, lossy, or lossless | Yes | High quality or smaller file size | Not universally viewable |
- Digital images are represented as rectangular arrays of square pixels.
- Digital images use a left-hand coordinate system, with the origin in the upper left corner, the x-axis running to the right, and the y-axis running down. Some learners may prefer to think in terms of counting down rows for the y-axis and across columns for the x-axis. Thus, we will make an effort to allow for both approaches in our lesson presentation.
- Most frequently, digital images use an additive RGB model, with eight bits for the red, green, and blue channels.
- scikit-image images are stored as multi-dimensional NumPy arrays.
- In scikit-image images, the red channel is specified first, then the green, then the blue, i.e., RGB.
- Lossless compression retains all the details in an image, but lossy compression results in loss of some of the original image detail.
- BMP images are uncompressed, meaning they have high quality but also that their file sizes are large.
- JPEG images use lossy compression, meaning that their file sizes are smaller, but image quality may suffer.
- TIFF images can be uncompressed or compressed with lossy or lossless compression.
- Depending on the camera or sensor, various useful pieces of information may be stored in an image file, in the image metadata.
Content from Reading Images
Last updated on 2025-10-15 | Edit this page
Overview
Questions
- Question 1
Objectives
- Find relevant documentation to load different proprietary file formats.
- Identify common proprietary microscopy file formats and understand how tools like BioIO support working with these formats.
- Extract relevant information from image metadata (channel names, stage positions, time frames…).
- Extract physical units from image metadata (ZYX and time) with BioIO.
- Manually estimate the size of an object in a bioimage in pixel size and physical units.
We have covered much of how images are represented in computer software. In this episode we will learn some more methods for accessing and changing digital images.
First, import the packages needed for this episode
Reading, displaying, and saving images
Imageio provides intuitive functions for reading and writing (saving) images. All of the popular image formats, such as BMP, PNG, JPEG, and TIFF are supported, along with several more esoteric formats. Check the Supported Formats docs for a list of all formats. Matplotlib provides a large collection of plotting utilities.
Let us examine a simple Python program to load, display, and save an image to a different format. Here are the first few lines:
PYTHON
"""Python program to open, display, and save an image."""
# read image
chair = iio.imread(uri="data/chair.jpg")We use the iio.imread() function to read a JPEG image
entitled chair.jpg. Imageio reads the image, converts
it from JPEG into a NumPy array, and returns the array; we save the
array in a variable named chair.
Next, we will do something with the image:
Once we have the image in the program, we first call
fig, ax = plt.subplots() so that we will have a fresh
figure with a set of axes independent from our previous calls. Next we
call ax.imshow() in order to display the image.
Now, we will save the image in another format:
The final statement in the program,
iio.imwrite(uri="data/chair.tif", image=chair), writes the
image to a file named chair.tif in the data/
directory. The imwrite() function automatically determines
the type of the file, based on the file extension we provide. In this
case, the .tif extension causes the image to be saved as a
TIFF.
Metadata, revisited
Remember, as mentioned in the previous section, images saved with
imwrite() will not retain all metadata associated with the
original image that was loaded into Python! If the image metadata
is important to you, be sure to always keep an unchanged copy of
the original image!
Extensions do not always dictate file type
The iio.imwrite() function automatically uses the file
type we specify in the file name parameter’s extension. Note that this
is not always the case. For example, if we are editing a document in
Microsoft Word, and we save the document as paper.pdf
instead of paper.docx, the file is not saved as a
PDF document.
Named versus positional arguments
When we call functions in Python, there are two ways we can specify the necessary arguments. We can specify the arguments positionally, i.e., in the order the parameters appear in the function definition, or we can use named arguments.
For example, the iio.imwrite() function
definition specifies two parameters, the resource to save the image
to (e.g., a file name, an http address) and the image to write to disk.
So, we could save the chair image in the sample code above using
positional arguments like this:
iio.imwrite("data/chair.tif", image)
Since the function expects the first argument to be the file name,
there is no confusion about what "data/chair.jpg" means.
The same goes for the second argument.
The style we will use in this workshop is to name each argument, like this:
iio.imwrite(uri="data/chair.tif", image=image)
This style will make it easier for you to learn how to use the variety of functions we will cover in this workshop.
Resizing an image (10 min)
Using the chair.jpg image located in the data folder,
write a Python script to read your image into a variable named
chair. Then, resize the image to 10 percent of its current
size using these lines of code:
PYTHON
new_shape = (chair.shape[0] // 10, chair.shape[1] // 10, chair.shape[2])
resized_chair = ski.transform.resize(image=chair, output_shape=new_shape)
resized_chair = ski.util.img_as_ubyte(resized_chair)As it is used here, the parameters to the
ski.transform.resize() function are the image to transform,
chair, the dimensions we want the new image to have,
new_shape.
Note that the pixel values in the new image are an approximation of
the original values and should not be confused with actual, observed
data. This is because scikit-image interpolates the pixel values when
reducing or increasing the size of an image.
ski.transform.resize has a number of optional parameters
that allow the user to control this interpolation. You can find more
details in the scikit-image
documentation.
Image files on disk are normally stored as whole numbers for space
efficiency, but transformations and other math operations often result
in conversion to floating point numbers. Using the
ski.util.img_as_ubyte() method converts it back to whole
numbers before we save it back to disk. If we don’t convert it before
saving, iio.imwrite() may not recognise it as image
data.
Next, write the resized image out to a new file named
resized.jpg in your data directory. Finally, use
ax.imshow() with each of your image variables to display
both images in your notebook. Don’t forget to use
fig, ax = plt.subplots() so you don’t overwrite the first
image with the second. Images may appear the same size in jupyter, but
you can see the size difference by comparing the scales for each. You
can also see the difference in file storage size on disk by hovering
your mouse cursor over the original and the new files in the Jupyter
file browser, using ls -l in your shell (dir
with Windows PowerShell), or viewing file sizes in the OS file browser
if it is configured so.
Here is what your Python script might look like.
PYTHON
"""Python script to read an image, resize it, and save it under a different name."""
# read in image
chair = iio.imread(uri="data/chair.jpg")
# resize the image
new_shape = (chair.shape[0] // 10, chair.shape[1] // 10, chair.shape[2])
resized_chair = ski.transform.resize(image=chair, output_shape=new_shape)
resized_chair = ski.util.img_as_ubyte(resized_chair)
# write out image
iio.imwrite(uri="data/resized_chair.jpg", image=resized_chair)
# display images
fig, ax = plt.subplots()
ax.imshow(chair)
fig, ax = plt.subplots()
ax.imshow(resized_chair)The script resizes the data/chair.jpg image by a factor
of 10 in both dimensions, saves the result to the
data/resized_chair.jpg file, and displays original and
resized for comparision.
Manipulating pixels
In the Image Basics episode, we individually manipulated the colours of pixels by changing the numbers stored in the image’s NumPy array. Let’s apply the principles learned there along with some new principles to a real world example.
Suppose we are interested in this maize root cluster image. We want to be able to focus our program’s attention on the roots themselves, while ignoring the black background.
Since the image is stored as an array of numbers, we can simply look through the array for pixel colour values that are less than some threshold value. This process is called thresholding, and we will see more powerful methods to perform the thresholding task in the Thresholding episode. Here, though, we will look at a simple and elegant NumPy method for thresholding. Let us develop a program that keeps only the pixel colour values in an image that have value greater than or equal to 128. This will keep the pixels that are brighter than half of “full brightness”, i.e., pixels that do not belong to the black background.
We will start by reading the image and displaying it.
Loading images with imageio: Read-only arrays
When loading an image with imageio, in certain situations the image
is stored in a read-only array. If you attempt to manipulate the pixels
in a read-only array, you will receive an error message
ValueError: assignment destination is read-only. In order
to make the image array writeable, we can create a copy with
image = np.array(image) before manipulating the pixel
values.
PYTHON
"""Python script to ignore low intensity pixels in an image."""
# read input image
maize_roots = iio.imread(uri="data/maize-root-cluster.jpg")
maize_roots = np.array(maize_roots)
# display original image
fig, ax = plt.subplots()
ax.imshow(maize_roots)Now we can threshold the image and display the result.
PYTHON
# keep only high-intensity pixels
maize_roots[maize_roots < 128] = 0
# display modified image
fig, ax = plt.subplots()
ax.imshow(maize_roots)The NumPy command to ignore all low-intensity pixels is
roots[roots < 128] = 0. Every pixel colour value in the
whole 3-dimensional array with a value less that 128 is set to zero. In
this case, the result is an image in which the extraneous background
detail has been removed.
Converting colour images to grayscale
It is often easier to work with grayscale images, which have a single
channel, instead of colour images, which have three channels.
scikit-image offers the function ski.color.rgb2gray() to
achieve this. This function adds up the three colour channels in a way
that matches human colour perception, see the
scikit-image documentation for details. It returns a grayscale image
with floating point values in the range from 0 to 1. We can use the
function ski.util.img_as_ubyte() in order to convert it
back to the original data type and the data range back 0 to 255. Note
that it is often better to use image values represented by floating
point values, because using floating point numbers is numerically more
stable.
Colour and color
The Carpentries generally prefers UK English spelling, which is why
we use “colour” in the explanatory text of this lesson. However,
scikit-image contains many modules and functions that include the US
English spelling, color. The exact spelling matters here,
e.g. you will encounter an error if you try to run
ski.colour.rgb2gray(). To account for this, we will use the
US English spelling, color, in example Python code
throughout the lesson. You will encounter a similar approach with
“centre” and center.
PYTHON
"""Python script to load a color image as grayscale."""
# read input image
chair = iio.imread(uri="data/chair.jpg")
# display original image
fig, ax = plt.subplots()
ax.imshow(chair)
# convert to grayscale and display
gray_chair = ski.color.rgb2gray(chair)
fig, ax = plt.subplots()
ax.imshow(gray_chair, cmap="gray")We can also load colour images as grayscale directly by passing the
argument mode="L" to iio.imread().
PYTHON
"""Python script to load a color image as grayscale."""
# read input image, based on filename parameter
gray_chair = iio.imread(uri="data/chair.jpg", mode="L")
# display grayscale image
fig, ax = plt.subplots()
ax.imshow(gray_chair, cmap="gray")The first argument to iio.imread() is the filename of
the image. The second argument mode="L" determines the type
and range of the pixel values in the image (e.g., an 8-bit pixel has a
range of 0-255). This argument is forwarded to the pillow
backend, a Python imaging library for which mode “L” means 8-bit pixels
and single-channel (i.e., grayscale). The backend used by
iio.imread() may be specified as an optional argument: to
use pillow, you would pass plugin="pillow". If
the backend is not specified explicitly, iio.imread()
determines the backend to use based on the image type.
Loading images with imageio: Pixel type and depth
When loading an image with mode="L", the pixel values
are stored as 8-bit integer numbers that can take values in the range
0-255. However, pixel values may also be stored with other types and
ranges. For example, some scikit-image functions return the pixel values
as floating point numbers in the range 0-1. The type and range of the
pixel values are important for the colorscale when plotting, and for
masking and thresholding images as we will see later in the lesson. If
you are unsure about the type of the pixel values, you can inspect it
with print(image.dtype). For the example above, you should
find that it is dtype('uint8') indicating 8-bit integer
numbers.
Keeping only low intensity pixels (10 min)
A little earlier, we showed how we could use Python and scikit-image to turn on only the high intensity pixels from an image, while turning all the low intensity pixels off. Now, you can practice doing the opposite - keeping all the low intensity pixels while changing the high intensity ones.
The file data/sudoku.png is an RGB image of a sudoku
puzzle:
Your task is to load the image in grayscale format and turn all of the bright pixels in the image to a light gray colour. In other words, mask the bright pixels that have a pixel value greater than, say, 192 and set their value to 192 (the value 192 is chosen here because it corresponds to 75% of the range 0-255 of an 8-bit pixel). The results should look like this:
Hint: the cmap, vmin, and
vmax parameters of matplotlib.pyplot.imshow
will be needed to display the modified image as desired. See the Matplotlib
documentation for more details on cmap,
vmin, and vmax.
First, load the image file data/sudoku.png as a
grayscale image. Note we may want to create a copy of the image array to
avoid modifying our original variable and also because
imageio.v3.imread sometimes returns a non-writeable
image.
PYTHON
sudoku = iio.imread(uri="data/sudoku.png", mode="L")
sudoku_gray_background = np.array(sudoku)Then change all bright pixel values greater than 192 to 192:
Finally, display the original and modified images side by side. Note
that we have to specify vmin=0 and vmax=255 as
the range of the colorscale because it would otherwise automatically
adjust to the new range 0-192.
Plotting single channel images (cmap, vmin, vmax)
Compared to a colour image, a grayscale image contains only a single
intensity value per pixel. When we plot such an image with
ax.imshow, Matplotlib uses a colour map, to assign each
intensity value a colour. The default colour map is called “viridis” and
maps low values to purple and high values to yellow. We can instruct
Matplotlib to map low values to black and high values to white instead,
by calling ax.imshow with cmap="gray". The
documentation contains an overview of pre-defined colour maps.
Furthermore, Matplotlib determines the minimum and maximum values of
the colour map dynamically from the image, by default. That means that
in an image where the minimum is 64 and the maximum is 192, those values
will be mapped to black and white respectively (and not dark gray and
light gray as you might expect). If there are defined minimum and
maximum vales, you can specify them via vmin and
vmax to get the desired output.
If you forget about this, it can lead to unexpected results. Try
removing the vmax parameter from the sudoku challenge
solution and see what happens.
Access via slicing
As noted in the previous lesson scikit-image images are stored as NumPy arrays, so we can use array slicing to select rectangular areas of an image. Then, we can save the selection as a new image, change the pixels in the image, and so on. It is important to remember that coordinates are specified in (ry, cx) order and that colour values are specified in (r, g, b) order when doing these manipulations.
Consider this image of a whiteboard, and suppose that we want to create a sub-image with just the portion that says “odd + even = odd,” along with the red box that is drawn around the words.
Using matplotlib.pyplot.imshow we can determine the
coordinates of the corners of the area we wish to extract by hovering
the mouse near the points of interest and noting the coordinates
(remember to run %matplotlib widget first if you haven’t
already). If we do that, we might settle on a rectangular area with an
upper-left coordinate of (135, 60) and a lower-right coordinate
of (480, 150), as shown in this version of the whiteboard
picture:
Note that the coordinates in the preceding image are specified in
(cx, ry) order. Now if our entire whiteboard image is stored as
a NumPy array named image, we can create a new image of the
selected region with a statement like this:
clip = image[60:151, 135:481, :]
Our array slicing specifies the range of y-coordinates or rows first,
60:151, and then the range of x-coordinates or columns,
135:481. Note we go one beyond the maximum value in each
dimension, so that the entire desired area is selected. The third part
of the slice, :, indicates that we want all three colour
channels in our new image.
A script to create the subimage would start by loading the image:
PYTHON
"""Python script demonstrating image modification and creation via NumPy array slicing."""
# load and display original image
board = iio.imread(uri="data/board.jpg")
board = np.array(board)
fig, ax = plt.subplots()
ax.imshow(board)Then we use array slicing to create a new image with our selected area and then display the new image.
PYTHON
# extract, display, and save sub-image
clipped_board = board[60:151, 135:481, :]
fig, ax = plt.subplots()
ax.imshow(clipped_board)
iio.imwrite(uri="data/clipped_board.tif", image=clipped_board)We can also change the values in an image, as shown next.
PYTHON
# replace clipped area with sampled color
color = board[330, 90]
board[60:151, 135:481] = color
fig, ax = plt.subplots()
ax.imshow(board)First, we sample a single pixel’s colour at a particular location of
the image, saving it in a variable named color, which
creates a 1 × 1 × 3 NumPy array with the blue, green, and red colour
values for the pixel located at (ry = 330, cx = 90). Then, with
the img[60:151, 135:481] = color command, we modify the
image in the specified area. From a NumPy perspective, this changes all
the pixel values within that range to array saved in the
color variable. In this case, the command “erases” that
area of the whiteboard, replacing the words with a beige colour, as
shown in the final image produced by the program:
Practicing with slices (10 min - optional, not included in timing)
Using the techniques you just learned, write a script that creates, displays, and saves a sub-image containing only the plant and its roots from “data/maize-root-cluster.jpg”
Here is the completed Python program to select only the plant and roots in the image.
PYTHON
"""Python script to extract a sub-image containing only the plant and roots in an existing image."""
# load and display original image
maize_roots = iio.imread(uri="data/maize-root-cluster.jpg")
fig, ax = plt.subplots()
ax.imshow(maize_roots)
# extract and display sub-image
clipped_maize = maize_roots[0:400, 275:550, :]
fig, ax = plt.subplots()
ax.imshow(clipped_maize)
# save sub-image
iio.imwrite(uri="data/clipped_maize.jpg", image=clipped_maize)- Images are read from disk with the
iio.imread()function. - We create a window that automatically scales the displayed image
with Matplotlib and calling
imshow()on the global figure object. - Colour images can be transformed to grayscale using
ski.color.rgb2gray()or, in many cases, be read as grayscale directly by passing the argumentmode="L"toiio.imread(). - We can resize images with the
ski.transform.resize()function. - NumPy array commands, such as
image[image < 128] = 0, can be used to manipulate the pixels of an image. - Array slicing can be used to extract sub-images or modify areas of
images, e.g.,
clip = image[60:150, 135:480, :]. - Metadata is not retained when images are loaded as NumPy arrays
using
iio.imread().
Content from Image exploration
Last updated on 2025-10-15 | Edit this page
Overview
Questions
- Question 1
Objectives
Create and visualize a histogram with Matplotlib.
Identify common problems with image quality by looking at histograms (saturation, clipping, dynamic range).
Explain the relationship between pixel intensity values, color maps (LUT) and physical fluorescence channels.
Describe how adjusting image display settings, including brightness, contrast, applying color maps, and windowing, affects the data.
Distinguish between linear and non-linear adjustments such as gamma correction.
Explain the importance of slicing, subsampling, and projections in image analysis.
Use image coordinates to access individual values and slices of NumPy arrays.
Select a subset of time frames of NumPy arrays using slicing.
Generate simple projections of NumPy arrays (max, time).
Visualize a single 2D NumPy array with Matplotlib.
Display images using different Matplotlib colormaps (i.e. LUTs) and contrast settings.
Display several images next to each other with Matplotlib subplots.
Add a colorbar to a displayed image with Matplotlib.
Open a Napari window and add one or more (multi-dimensional) images to it.
This is a new episode that didn’t exist in the original curriculum
It could take content both from 04-drawing.md and
05-creating-histograms.md.
Content from Preprocessing and filtering
Last updated on 2025-10-15 | Edit this page
Overview
Questions
- Question 1
Objectives
Apply histogram normalization and equalization to facilitate downstream image segmentation
Choose when normalization and equalization is appropriate (batch normalization of intensities)
Explain the concept of a filter and its effects on the image data (theoretical explanation of kernels)
Define the concept of noise and background in images
Apply techniques like gaussian, mean or median filtering to remove noise from an image
Use edge filter (or DoG/LoG filter) to highlight object edges (or spots) in an image
This is a new episode that didn’t exist in the original curriculum
It could take content both from
05-creating-histograms.md and
06-blurring.md.
Content from Processing and segmentation
Last updated on 2025-10-15 | Edit this page
Overview
Questions
- How can we use thresholding to produce a binary image?
Objectives
Choose between semantic and instance segmentation, and detection, according to the goals of an analysis
Compare manual and algorithmic thresholding techniques with their implications for reproducibility
Identify connected pixels in images to identify objects of interest
Apply labeling techniques using connected components
Visualize labeled objects as an overlay of raw data
In this episode, we will learn how to use scikit-image functions to apply thresholding to an image. Thresholding is a type of image segmentation, where we change the pixels of an image to make the image easier to analyze. In thresholding, we convert an image from colour or grayscale into a binary image, i.e., one that is simply black and white. Most frequently, we use thresholding as a way to select areas of interest of an image, while ignoring the parts we are not concerned with. We have already done some simple thresholding, in the “Manipulating pixels” section of the Working with scikit-image episode. In that case, we used a simple NumPy array manipulation to separate the pixels belonging to the root system of a plant from the black background. In this episode, we will learn how to use scikit-image functions to perform thresholding. Then, we will use the masks returned by these functions to select the parts of an image we are interested in.
First, import the packages needed for this episode
Simple thresholding
Consider the image data/shapes-01.jpg with a series of
crudely cut shapes set against a white background.
PYTHON
# load the image
shapes01 = iio.imread(uri="data/shapes-01.jpg")
fig, ax = plt.subplots()
ax.imshow(shapes01)
Now suppose we want to select only the shapes from the image. In
other words, we want to leave the pixels belonging to the shapes “on,”
while turning the rest of the pixels “off,” by setting their colour
channel values to zeros. The scikit-image library has several different
methods of thresholding. We will start with the simplest version, which
involves an important step of human input. Specifically, in this simple,
fixed-level thresholding, we have to provide a threshold value
t.
The process works like this. First, we will load the original image, convert it to grayscale, and de-noise it as in the Blurring Images episode.
PYTHON
# convert the image to grayscale
gray_shapes = ski.color.rgb2gray(shapes01)
# blur the image to denoise
blurred_shapes = ski.filters.gaussian(gray_shapes, sigma=1.0)
fig, ax = plt.subplots()
ax.imshow(blurred_shapes, cmap="gray")
Next, we would like to apply the threshold t such that
pixels with grayscale values on one side of t will be
turned “on”, while pixels with grayscale values on the other side will
be turned “off”. How might we do that? Remember that grayscale images
contain pixel values in the range from 0 to 1, so we are looking for a
threshold t in the closed range [0.0, 1.0]. We see in the
image that the geometric shapes are “darker” than the white background
but there is also some light gray noise on the background. One way to
determine a “good” value for t is to look at the grayscale
histogram of the image and try to identify what grayscale ranges
correspond to the shapes in the image or the background.
The histogram for the shapes image shown above can be produced as in the Creating Histograms episode.
PYTHON
# create a histogram of the blurred grayscale image
histogram, bin_edges = np.histogram(blurred_shapes, bins=256, range=(0.0, 1.0))
fig, ax = plt.subplots()
ax.plot(bin_edges[0:-1], histogram)
ax.set_title("Grayscale Histogram")
ax.set_xlabel("grayscale value")
ax.set_ylabel("pixels")
ax.set_xlim(0, 1.0)
Since the image has a white background, most of the pixels in the
image are white. This corresponds nicely to what we see in the
histogram: there is a peak near the value of 1.0. If we want to select
the shapes and not the background, we want to turn off the white
background pixels, while leaving the pixels for the shapes turned on.
So, we should choose a value of t somewhere before the
large peak and turn pixels above that value “off”. Let us choose
t=0.8.
To apply the threshold t, we can use the NumPy
comparison operators to create a mask. Here, we want to turn “on” all
pixels which have values smaller than the threshold, so we use the less
operator < to compare the blurred_image to
the threshold t. The operator returns a mask, that we
capture in the variable binary_mask. It has only one
channel, and each of its values is either 0 or 1. The binary mask
created by the thresholding operation can be shown with
ax.imshow, where the False entries are shown
as black pixels (0-valued) and the True entries are shown
as white pixels (1-valued).
PYTHON
# create a mask based on the threshold
t = 0.8
binary_mask = blurred_shapes < t
fig, ax = plt.subplots()
ax.imshow(binary_mask, cmap="gray")
You can see that the areas where the shapes were in the original area are now white, while the rest of the mask image is black.
What makes a good threshold?
As is often the case, the answer to this question is “it depends”. In
the example above, we could have just switched off all the white
background pixels by choosing t=1.0, but this would leave
us with some background noise in the mask image. On the other hand, if
we choose too low a value for the threshold, we could lose some of the
shapes that are too bright. You can experiment with the threshold by
re-running the above code lines with different values for
t. In practice, it is a matter of domain knowledge and
experience to interpret the peaks in the histogram so to determine an
appropriate threshold. The process often involves trial and error, which
is a drawback of the simple thresholding method. Below we will introduce
automatic thresholding, which uses a quantitative, mathematical
definition for a good threshold that allows us to determine the value of
t automatically. It is worth noting that the principle for
simple and automatic thresholding can also be used for images with pixel
ranges other than [0.0, 1.0]. For example, we could perform thresholding
on pixel intensity values in the range [0, 255] as we have already seen
in the Working with
scikit-image episode.
We can now apply the binary_mask to the original
coloured image as we have learned in the
Drawing and Bitwise Operations episode. What we are left
with is only the coloured shapes from the original.
PYTHON
# use the binary_mask to select the "interesting" part of the image
selection = shapes01.copy()
selection[~binary_mask] = 0
fig, ax = plt.subplots()
ax.imshow(selection)
More practice with simple thresholding (15 min)
Now, it is your turn to practice. Suppose we want to use simple
thresholding to select only the coloured shapes (in this particular case
we consider grayish to be a colour, too) from the image
data/shapes-02.jpg:
First, plot the grayscale histogram as in the Creating Histogram episode and
examine the distribution of grayscale values in the image. What do you
think would be a good value for the threshold t?
The histogram for the data/shapes-02.jpg image can be
shown with
PYTHON
shapes = iio.imread(uri="data/shapes-02.jpg")
gray_shapes = ski.color.rgb2gray(shapes)
histogram, bin_edges = np.histogram(gray_shapes, bins=256, range=(0.0, 1.0))
fig, ax = plt.subplots()
ax.plot(bin_edges[0:-1], histogram)
ax.set_title("Graylevel histogram")
ax.set_xlabel("gray value")
ax.set_ylabel("pixel count")
ax.set_xlim(0, 1.0)
We can see a large spike around 0.3, and a smaller spike around 0.7.
The spike near 0.3 represents the darker background, so it seems like a
value close to t=0.5 would be a good choice.
More practice with simple thresholding (15 min) (continued)
Next, create a mask to turn the pixels above the threshold
t on and pixels below the threshold t off.
Note that unlike the image with a white background we used above, here
the peak for the background colour is at a lower gray level than the
shapes. Therefore, change the comparison operator less <
to greater > to create the appropriate mask. Then apply
the mask to the image and view the thresholded image. If everything
works as it should, your output should show only the coloured shapes on
a black background.
Here are the commands to create and view the binary mask
PYTHON
t = 0.5
binary_mask = gray_shapes > t
fig, ax = plt.subplots()
ax.imshow(binary_mask, cmap="gray")
And here are the commands to apply the mask and view the thresholded image
PYTHON
shapes02 = iio.imread(uri="data/shapes-02.jpg")
selection = shapes02.copy()
selection[~binary_mask] = 0
fig, ax = plt.subplots()
ax.imshow(selection)
Automatic thresholding
The downside of the simple thresholding technique is that we have to
make an educated guess about the threshold t by inspecting
the histogram. There are also automatic thresholding methods
that can determine the threshold automatically for us. One such method
is Otsu’s
method. It is particularly useful for situations where the
grayscale histogram of an image has two peaks that correspond to
background and objects of interest.
Denoising an image before thresholding
In practice, it is often necessary to denoise the image before thresholding, which can be done with one of the methods from the Blurring Images episode.
Consider the image data/maize-root-cluster.jpg of a
maize root system which we have seen before in the Working with scikit-image
episode.
PYTHON
maize_roots = iio.imread(uri="data/maize-root-cluster.jpg")
fig, ax = plt.subplots()
ax.imshow(maize_roots)We use Gaussian blur with a sigma of 1.0 to denoise the root image. Let us look at the grayscale histogram of the denoised image.
PYTHON
# convert the image to grayscale
gray_image = ski.color.rgb2gray(maize_roots)
# blur the image to denoise
blurred_image = ski.filters.gaussian(gray_image, sigma=1.0)
# show the histogram of the blurred image
histogram, bin_edges = np.histogram(blurred_image, bins=256, range=(0.0, 1.0))
fig, ax = plt.subplots()
ax.plot(bin_edges[0:-1], histogram)
ax.set_title("Graylevel histogram")
ax.set_xlabel("gray value")
ax.set_ylabel("pixel count")
ax.set_xlim(0, 1.0)
The histogram has a significant peak around 0.2 and then a broader “hill” around 0.6 followed by a smaller peak near 1.0. Looking at the grayscale image, we can identify the peak at 0.2 with the background and the broader peak with the foreground. Thus, this image is a good candidate for thresholding with Otsu’s method. The mathematical details of how this works are complicated (see the scikit-image documentation if you are interested), but the outcome is that Otsu’s method finds a threshold value between the two peaks of a grayscale histogram which might correspond well to the foreground and background depending on the data and application.
The ski.filters.threshold_otsu() function can be used to
determine the threshold automatically via Otsu’s method. Then NumPy
comparison operators can be used to apply it as before. Here are the
Python commands to determine the threshold t with Otsu’s
method.
PYTHON
# perform automatic thresholding
t = ski.filters.threshold_otsu(blurred_image)
print("Found automatic threshold t = {}.".format(t))OUTPUT
Found automatic threshold t = 0.4172454549881862.For this root image and a Gaussian blur with the chosen sigma of 1.0,
the computed threshold value is 0.42. No we can create a binary mask
with the comparison operator >. As we have seen before,
pixels above the threshold value will be turned on, those below the
threshold will be turned off.
PYTHON
# create a binary mask with the threshold found by Otsu's method
binary_mask = blurred_image > t
fig, ax = plt.subplots()
ax.imshow(binary_mask, cmap="gray")
Finally, we use the mask to select the foreground:
PYTHON
# apply the binary mask to select the foreground
selection = maize_roots.copy()
selection[~binary_mask] = 0
fig, ax = plt.subplots()
ax.imshow(selection)
Application: measuring root mass
Let us now turn to an application where we can apply thresholding and
other techniques we have learned to this point. Consider these four
maize root system images, which you can find in the files
data/trial-016.jpg, data/trial-020.jpg,
data/trial-216.jpg, and
data/trial-293.jpg.
Suppose we are interested in the amount of plant material in each image, and in particular how that amount changes from image to image. Perhaps the images represent the growth of the plant over time, or perhaps the images show four different maize varieties at the same phase of their growth. The question we would like to answer is, “how much root mass is in each image?”
We will first construct a Python program to measure this value for a single image. Our strategy will be this:
- Read the image, converting it to grayscale as it is read. For this application we do not need the colour image.
- Blur the image.
- Use Otsu’s method of thresholding to create a binary image, where the pixels that were part of the maize plant are white, and everything else is black.
- Save the binary image so it can be examined later.
- Count the white pixels in the binary image, and divide by the number of pixels in the image. This ratio will be a measure of the root mass of the plant in the image.
- Output the name of the image processed and the root mass ratio.
Our intent is to perform these steps and produce the numeric result - a measure of the root mass in the image - without human intervention. Implementing the steps within a Python function will enable us to call this function for different images.
Here is a Python function that implements this root-mass-measuring strategy. Since the function is intended to produce numeric output without human interaction, it does not display any of the images. Almost all of the commands should be familiar, and in fact, it may seem simpler than the code we have worked on thus far, because we are not displaying any of the images.
PYTHON
def measure_root_mass(filename, sigma=1.0):
# read the original image, converting to grayscale on the fly
image = iio.imread(uri=filename, mode="L")
# blur before thresholding
blurred_image = ski.filters.gaussian(image, sigma=sigma)
# perform automatic thresholding to produce a binary image
t = ski.filters.threshold_otsu(blurred_image)
binary_mask = blurred_image > t
# determine root mass ratio
root_pixels = np.count_nonzero(binary_mask)
density = root_pixels / binary_mask.size
return densityThe function begins with reading the original image from the file
filename. We use iio.imread() with the
optional argument mode="L" to automatically convert it to
grayscale. Next, the grayscale image is blurred with a Gaussian filter
with the value of sigma that is passed to the function.
Then we determine the threshold t with Otsu’s method and
create a binary mask just as we did in the previous section. Up to this
point, everything should be familiar.
The final part of the function determines the root mass ratio in the
image. Recall that in the binary_mask, every pixel has
either a value of zero (black/background) or one (white/foreground). We
want to count the number of white pixels, which can be accomplished with
a call to the NumPy function np.count_nonzero. Finally, the
density ratio is calculated by dividing the number of white pixels by
the total number of pixels binary_mask.size in the image.
The function returns then root density of the image.
We can call this function with any filename and provide a sigma value
for the blurring. If no sigma value is provided, the default value 1.0
will be used. For example, for the file data/trial-016.jpg
and a sigma value of 1.5, we would call the function like this:
OUTPUT
0.0482436835106383`Now we can use the function to process the series of four images shown above. In a real-world scientific situation, there might be dozens, hundreds, or even thousands of images to process. To save us the tedium of calling the function for each image by hand, we can write a loop that processes all files automatically. The following code block assumes that the files are located in the same directory and the filenames all start with the trial- prefix and end with the .jpg suffix.
PYTHON
all_files = glob.glob("data/trial-*.jpg")
for filename in all_files:
density = measure_root_mass(filename=filename, sigma=1.5)
# output in format suitable for .csv
print(filename, density, sep=",")OUTPUT
data/trial-016.jpg,0.0482436835106383
data/trial-020.jpg,0.06346941489361702
data/trial-216.jpg,0.14073969414893617
data/trial-293.jpg,0.13607895611702128Ignoring more of the images – brainstorming (10 min)
Let us take a closer look at the binary masks produced by the
measure_root_mass function.
You may have noticed in the section on automatic thresholding that the thresholded image does include regions of the image aside of the plant root: the numbered labels and the white circles in each image are preserved during the thresholding, because their grayscale values are above the threshold. Therefore, our calculated root mass ratios include the white pixels of the label and white circle that are not part of the plant root. Those extra pixels affect how accurate the root mass calculation is!
How might we remove the labels and circles before calculating the ratio, so that our results are more accurate? Think about some options given what we have learned so far.
One approach we might take is to try to completely mask out a region from each image, particularly, the area containing the white circle and the numbered label. If we had coordinates for a rectangular area on the image that contained the circle and the label, we could mask the area out by using techniques we learned in the Drawing and Bitwise Operations episode.
However, a closer inspection of the binary images raises some issues with that approach. Since the roots are not always constrained to a certain area in the image, and since the circles and labels are in different locations each time, we would have difficulties coming up with a single rectangle that would work for every image. We could create a different masking rectangle for each image, but that is not a practicable approach if we have hundreds or thousands of images to process.
Another approach we could take is to apply two thresholding steps to
the image. Look at the graylevel histogram of the file
data/trial-016.jpg shown above again: Notice the peak near
1.0? Recall that a grayscale value of 1.0 corresponds to white pixels:
the peak corresponds to the white label and circle. So, we could use
simple binary thresholding to mask the white circle and label from the
image, and then we could use Otsu’s method to select the pixels in the
plant portion of the image.
Note that most of this extra work in processing the image could have been avoided during the experimental design stage, with some careful consideration of how the resulting images would be used. For example, all of the following measures could have made the images easier to process, by helping us predict and/or detect where the label is in the image and subsequently mask it from further processing:
- Using labels with a consistent size and shape
- Placing all the labels in the same position, relative to the sample
- Using a non-white label, with non-black writing
Ignoring more of the images – implementation (30 min - optional, not included in timing)
Implement an enhanced version of the function
measure_root_mass that applies simple binary thresholding
to remove the white circle and label from the image before applying
Otsu’s method.
We can apply a simple binary thresholding with a threshold
t=0.95 to remove the label and circle from the image. We
can then use the binary mask to calculate the Otsu threshold without the
pixels from the label and circle.
PYTHON
def enhanced_root_mass(filename, sigma):
# read the original image, converting to grayscale on the fly
image = iio.imread(uri=filename, mode="L")
# blur before thresholding
blurred_image = ski.filters.gaussian(image, sigma=sigma)
# perform binary thresholding to mask the white label and circle
binary_mask = blurred_image < 0.95
# perform automatic thresholding using only the pixels with value True in the binary mask
t = ski.filters.threshold_otsu(blurred_image[binary_mask])
# update binary mask to identify pixels which are both less than 0.95 and greater than t
binary_mask = (blurred_image < 0.95) & (blurred_image > t)
# determine root mass ratio
root_pixels = np.count_nonzero(binary_mask)
density = root_pixels / binary_mask.size
return density
all_files = glob.glob("data/trial-*.jpg")
for filename in all_files:
density = enhanced_root_mass(filename=filename, sigma=1.5)
# output in format suitable for .csv
print(filename, density, sep=",")The output of the improved program does illustrate that the white circles and labels were skewing our root mass ratios:
OUTPUT
data/trial-016.jpg,0.046250166223404256
data/trial-020.jpg,0.05886968085106383
data/trial-216.jpg,0.13712117686170214
data/trial-293.jpg,0.13190342420212767The & operator above means that we have defined a
logical AND statement. This combines the two tests of pixel intensities
in the blurred image such that both must be true for a pixel’s position
to be set to True in the resulting mask.
Result of t < blurred_image
|
Result of blurred_image < 0.95
|
Resulting value in binary_mask
|
|---|---|---|
| False | True | False |
| True | False | False |
| True | True | True |
Knowing how to construct this kind of logical operation can be very helpful in image processing. The University of Minnesota Library’s guide to Boolean operators is a good place to start if you want to learn more.
Here are the binary images produced by the additional thresholding. Note that we have not completely removed the offending white pixels. Outlines still remain. However, we have reduced the number of extraneous pixels, which should make the output more accurate.
Thresholding a bacteria colony image (15 min)
In the images directory data/, you will find an image
named colonies-01.tif.
This is one of the images you will be working with in the morphometric challenge at the end of the workshop.
- Plot and inspect the grayscale histogram of the image to determine a good threshold value for the image.
- Create a binary mask that leaves the pixels in the bacteria colonies “on” while turning the rest of the pixels in the image “off”.
Here is the code to create the grayscale histogram:
PYTHON
bacteria = iio.imread(uri="data/colonies-01.tif")
gray_image = ski.color.rgb2gray(bacteria)
blurred_image = ski.filters.gaussian(gray_image, sigma=1.0)
histogram, bin_edges = np.histogram(blurred_image, bins=256, range=(0.0, 1.0))
fig, ax = plt.subplots()
ax.plot(bin_edges[0:-1], histogram)
ax.set_title("Graylevel histogram")
ax.set_xlabel("gray value")
ax.set_ylabel("pixel count")
ax.set_xlim(0, 1.0)
The peak near one corresponds to the white image background, and the
broader peak around 0.5 corresponds to the yellow/brown culture medium
in the dish. The small peak near zero is what we are after: the dark
bacteria colonies. A reasonable choice thus might be to leave pixels
below t=0.2 on.
Here is the code to create and show the binarized image using the
< operator with a threshold t=0.2:
PYTHON
t = 0.2
binary_mask = blurred_image < t
fig, ax = plt.subplots()
ax.imshow(binary_mask, cmap="gray")
When you experiment with the threshold a bit, you can see that in particular the size of the bacteria colony near the edge of the dish in the top right is affected by the choice of the threshold.
- Thresholding produces a binary image, where all pixels with intensities above (or below) a threshold value are turned on, while all other pixels are turned off.
- The binary images produced by thresholding are held in two-dimensional NumPy arrays, since they have only one colour value channel. They are boolean, hence they contain the values 0 (off) and 1 (on).
- Thresholding can be used to create masks that select only the interesting parts of an image, or as the first step before edge detection or finding contours.
Content from Post-processing with filters
Last updated on 2025-10-15 | Edit this page
Overview
Questions
- Question 1
Objectives
- Utilize post-processing filters such as watershed, fill holes, opening/closing techniques, and removing objects touching image borders to improve objects segmentation for later measurements
Objects
In the Thresholding episode we have covered dividing an image into foreground and background pixels. In the shapes example image, we considered the coloured shapes as foreground objects on a white background.
In thresholding we went from the original image to this version:
Here, we created a mask that only highlights the parts of the image
that we find interesting, the objects. All objects have pixel
value of True while the background pixels are
False.
By looking at the mask image, one can count the objects that are present in the image (7). But how did we actually do that, how did we decide which lump of pixels constitutes a single object?
Pixel Neighborhoods
In order to decide which pixels belong to the same object, one can exploit their neighborhood: pixels that are directly next to each other and belong to the foreground class can be considered to belong to the same object.
Let’s discuss the concept of pixel neighborhoods in more detail.
Consider the following mask “image” with 8 rows, and 8 columns. For the
purpose of illustration, the digit 0 is used to represent
background pixels, and the letter X is used to represent
object pixels foreground).
OUTPUT
0 0 0 0 0 0 0 0
0 X X 0 0 0 0 0
0 X X 0 0 0 0 0
0 0 0 X X X 0 0
0 0 0 X X X X 0
0 0 0 0 0 0 0 0The pixels are organised in a rectangular grid. In order to
understand pixel neighborhoods we will introduce the concept of “jumps”
between pixels. The jumps follow two rules: First rule is that one jump
is only allowed along the column, or the row. Diagonal jumps are not
allowed. So, from a centre pixel, denoted with o, only the
pixels indicated with a 1 are reachable:
OUTPUT
- 1 -
1 o 1
- 1 -The pixels on the diagonal (from o) are not reachable
with a single jump, which is denoted by the -. The pixels
reachable with a single jump form the 1-jump
neighborhood.
The second rule states that in a sequence of jumps, one may only jump
in row and column direction once -> they have to be
orthogonal. An example of a sequence of orthogonal jumps is
shown below. Starting from o the first jump goes along the
row to the right. The second jump then goes along the column direction
up. After this, the sequence cannot be continued as a jump has already
been made in both row and column direction.
OUTPUT
- - 2
- o 1
- - -All pixels reachable with one, or two jumps form the
2-jump neighborhood. The grid below illustrates the
pixels reachable from the centre pixel o with a single
jump, highlighted with a 1, and the pixels reachable with 2
jumps with a 2.
OUTPUT
2 1 2
1 o 1
2 1 2We want to revisit our example image mask from above and apply the
two different neighborhood rules. With a single jump connectivity for
each pixel, we get two resulting objects, highlighted in the image with
A’s and B’s.
OUTPUT
0 0 0 0 0 0 0 0
0 A A 0 0 0 0 0
0 A A 0 0 0 0 0
0 0 0 B B B 0 0
0 0 0 B B B B 0
0 0 0 0 0 0 0 0In the 1-jump version, only pixels that have direct neighbors along
rows or columns are considered connected. Diagonal connections are not
included in the 1-jump neighborhood. With two jumps, however, we only
get a single object A because pixels are also considered
connected along the diagonals.
OUTPUT
0 0 0 0 0 0 0 0
0 A A 0 0 0 0 0
0 A A 0 0 0 0 0
0 0 0 A A A 0 0
0 0 0 A A A A 0
0 0 0 0 0 0 0 0Object counting (optional, not included in timing)
How many objects with 1 orthogonal jump, how many with 2 orthogonal jumps?
OUTPUT
0 0 0 0 0 0 0 0
0 X 0 0 0 X X 0
0 0 X 0 0 0 0 0
0 X 0 X X X 0 0
0 X 0 X X 0 0 0
0 0 0 0 0 0 0 01 jump
- 1
- 5
- 2
- 5
Object counting (optional, not included in timing) (continued)
2 jumps
- 2
- 3
- 5
- 2
Jumps and neighborhoods
We have just introduced how you can reach different neighboring pixels by performing one or more orthogonal jumps. We have used the terms 1-jump and 2-jump neighborhood. There is also a different way of referring to these neighborhoods: the 4- and 8-neighborhood. With a single jump you can reach four pixels from a given starting pixel. Hence, the 1-jump neighborhood corresponds to the 4-neighborhood. When two orthogonal jumps are allowed, eight pixels can be reached, so the 2-jump neighborhood corresponds to the 8-neighborhood.
Connected Component Analysis
In order to find the objects in an image, we want to employ an
operation that is called Connected Component Analysis (CCA). This
operation takes a binary image as an input. Usually, the
False value in this image is associated with background
pixels, and the True value indicates foreground, or object
pixels. Such an image can be produced, e.g., with thresholding. Given a
thresholded image, the connected component analysis produces a new
labeled image with integer pixel values. Pixels with the same
value, belong to the same object. scikit-image provides connected
component analysis in the function ski.measure.label(). Let
us add this function to the already familiar steps of thresholding an
image.
First, import the packages needed for this episode:
PYTHON
import imageio.v3 as iio
import ipympl
import matplotlib.pyplot as plt
import numpy as np
import skimage as ski
%matplotlib widgetIn this episode, we will use the ski.measure.label
function to perform the CCA.
Next, we define a reusable Python function
connected_components:
PYTHON
def connected_components(filename, sigma=1.0, t=0.5, connectivity=2):
# load the image
image = iio.imread(filename)
# convert the image to grayscale
gray_image = ski.color.rgb2gray(image)
# denoise the image with a Gaussian filter
blurred_image = ski.filters.gaussian(gray_image, sigma=sigma)
# mask the image according to threshold
binary_mask = blurred_image < t
# perform connected component analysis
labeled_image, count = ski.measure.label(binary_mask,
connectivity=connectivity, return_num=True)
return labeled_image, countThe first four lines of code are familiar from the Thresholding episode.
Then we call the ski.measure.label function. This
function has one positional argument where we pass the
binary_mask, i.e., the binary image to work on. With the
optional argument connectivity, we specify the neighborhood
in units of orthogonal jumps. For example, by setting
connectivity=2 we will consider the 2-jump neighborhood
introduced above. The function returns a labeled_image
where each pixel has a unique value corresponding to the object it
belongs to. In addition, we pass the optional parameter
return_num=True to return the maximum label index as
count.
Optional parameters and return values
The optional parameter return_num changes the data type
that is returned by the function ski.measure.label. The
number of labels is only returned if return_num is
True. Otherwise, the function only returns the labeled image.
This means that we have to pay attention when assigning the return value
to a variable. If we omit the optional parameter return_num
or pass return_num=False, we can call the function as
If we pass return_num=True, the function returns a tuple
and we can assign it as
If we used the same assignment as in the first case, the variable
labeled_image would become a tuple, in which
labeled_image[0] is the image and
labeled_image[1] is the number of labels. This could cause
confusion if we assume that labeled_image only contains the
image and pass it to other functions. If you get an
AttributeError: 'tuple' object has no attribute 'shape' or
similar, check if you have assigned the return values consistently with
the optional parameters.
We can call the above function connected_components and
display the labeled image like so:
PYTHON
labeled_image, count = connected_components(filename="data/shapes-01.jpg", sigma=2.0, t=0.9, connectivity=2)
fig, ax = plt.subplots()
ax.imshow(labeled_image)
ax.set_axis_off();If you are using an older version of Matplotlib you might get a
warning
UserWarning: Low image data range; displaying image with stretched contrast.
or just see a visually empty image.
What went wrong? When you hover over the image, the pixel values are
shown as numbers in the lower corner of the viewer. You can see that
some pixels have values different from 0, so they are not
actually all the same value. Let’s find out more by examining
labeled_image. Properties that might be interesting in this
context are dtype, the minimum and maximum value. We can
print them with the following lines:
PYTHON
print("dtype:", labeled_image.dtype)
print("min:", np.min(labeled_image))
print("max:", np.max(labeled_image))Examining the output can give us a clue why the image appears empty.
OUTPUT
dtype: int32
min: 0
max: 11The dtype of labeled_image is
int32. This means that values in this image range from
-2 ** 31 to 2 ** 31 - 1. Those are really big
numbers. From this available space we only use the range from
0 to 11. When showing this image in the
viewer, it may squeeze the complete range into 256 gray values.
Therefore, the range of our numbers does not produce any visible
variation. One way to rectify this is to explicitly specify the data
range we want the colormap to cover:
PYTHON
fig, ax = plt.subplots()
ax.imshow(labeled_image, vmin=np.min(labeled_image), vmax=np.max(labeled_image))Note this is the default behaviour for newer versions of
matplotlib.pyplot.imshow. Alternatively we could convert
the image to RGB and then display it.
Suppressing outputs in Jupyter Notebooks
We just used ax.set_axis_off(); to hide the axis from
the image for a visually cleaner figure. The semicolon is added to
supress the output(s) of the statement, in this case
the axis limits. This is specific to Jupyter Notebooks.
We can use the function ski.color.label2rgb() to convert
the 32-bit grayscale labeled image to standard RGB colour (recall that
we already used the ski.color.rgb2gray() function to
convert to grayscale). With ski.color.label2rgb(), all
objects are coloured according to a list of colours that can be
customised. We can use the following commands to convert and show the
image:
PYTHON
# convert the label image to color image
colored_label_image = ski.color.label2rgb(labeled_image, bg_label=0)
fig, ax = plt.subplots()
ax.imshow(colored_label_image)
ax.set_axis_off();
How many objects are in that image (15 min)
Now, it is your turn to practice. Using the function
connected_components, find two ways of printing out the
number of objects found in the image.
What number of objects would you expect to get?
How does changing the sigma and threshold
values influence the result?
As you might have guessed, the return value count
already contains the number of objects found in the image. So it can
simply be printed with
But there is also a way to obtain the number of found objects from
the labeled image itself. Recall that all pixels that belong to a single
object are assigned the same integer value. The connected component
algorithm produces consecutive numbers. The background gets the value
0, the first object gets the value 1, the
second object the value 2, and so on. This means that by
finding the object with the maximum value, we also know how many objects
there are in the image. We can thus use the np.max function
from NumPy to find the maximum value that equals the number of found
objects:
Invoking the function with sigma=2.0, and
threshold=0.9, both methods will print
OUTPUT
Found 11 objects in the image.Lowering the threshold will result in fewer objects. The higher the threshold is set, the more objects are found. More and more background noise gets picked up as objects. Larger sigmas produce binary masks with less noise and hence a smaller number of objects. Setting sigma too high bears the danger of merging objects.
You might wonder why the connected component analysis with
sigma=2.0, and threshold=0.9 finds 11 objects,
whereas we would expect only 7 objects. Where are the four additional
objects? With a bit of detective work, we can spot some small objects in
the image, for example, near the left border.
For us it is clear that these small spots are artifacts and not
objects we are interested in. But how can we tell the computer? One way
to calibrate the algorithm is to adjust the parameters for blurring
(sigma) and thresholding (t), but you may have
noticed during the above exercise that it is quite hard to find a
combination that produces the right output number. In some cases,
background noise gets picked up as an object. And with other parameters,
some of the foreground objects get broken up or disappear completely.
Therefore, we need other criteria to describe desired properties of the
objects that are found.
Morphometrics - Describe object features with numbers
Morphometrics is concerned with the quantitative analysis of objects
and considers properties such as size and shape. For the example of the
images with the shapes, our intuition tells us that the objects should
be of a certain size or area. So we could use a minimum area as a
criterion for when an object should be detected. To apply such a
criterion, we need a way to calculate the area of objects found by
connected components. Recall how we determined the root mass in the Thresholding episode by
counting the pixels in the binary mask. But here we want to calculate
the area of several objects in the labeled image. The scikit-image
library provides the function ski.measure.regionprops to
measure the properties of labeled regions. It returns a list of
RegionProperties that describe each connected region in the
images. The properties can be accessed using the attributes of the
RegionProperties data type. Here we will use the properties
"area" and "label". You can explore the
scikit-image documentation to learn about other properties
available.
We can get a list of areas of the labeled objects as follows:
PYTHON
# compute object features and extract object areas
object_features = ski.measure.regionprops(labeled_image)
object_areas = [objf["area"] for objf in object_features]
object_areasThis will produce the output
OUTPUT
[318542, 1, 523204, 496613, 517331, 143, 256215, 1, 68, 338784, 265755]Plot a histogram of the object area distribution (10 min)
Similar to how we determined a “good” threshold in the Thresholding episode, it is often helpful to inspect the histogram of an object property. For example, we want to look at the distribution of the object areas.
- Create and examine a histogram of the object areas
obtained with
ski.measure.regionprops. - What does the histogram tell you about the objects?
The histogram can be plotted with
PYTHON
fig, ax = plt.subplots()
ax.hist(object_areas)
ax.set_xlabel("Area (pixels)")
ax.set_ylabel("Number of objects");
The histogram shows the number of objects (vertical axis) whose area is within a certain range (horizontal axis). The height of the bars in the histogram indicates the prevalence of objects with a certain area. The whole histogram tells us about the distribution of object sizes in the image. It is often possible to identify gaps between groups of bars (or peaks if we draw the histogram as a continuous curve) that tell us about certain groups in the image.
In this example, we can see that there are four small objects that
contain less than 50000 pixels. Then there is a group of four (1+1+2)
objects in the range between 200000 and 400000, and three objects with a
size around 500000. For our object count, we might want to disregard the
small objects as artifacts, i.e, we want to ignore the leftmost bar of
the histogram. We could use a threshold of 50000 as the minimum area to
count. In fact, the object_areas list already tells us that
there are fewer than 200 pixels in these objects. Therefore, it is
reasonable to require a minimum area of at least 200 pixels for a
detected object. In practice, finding the “right” threshold can be
tricky and usually involves an educated guess based on domain
knowledge.
Filter objects by area (10 min)
Now we would like to use a minimum area criterion to obtain a more accurate count of the objects in the image.
- Find a way to calculate the number of objects by only counting objects above a certain area.
One way to count only objects above a certain area is to first create a list of those objects, and then take the length of that list as the object count. This can be done as follows:
PYTHON
min_area = 200
large_objects = []
for objf in object_features:
if objf["area"] > min_area:
large_objects.append(objf["label"])
print("Found", len(large_objects), "objects!")Another option is to use NumPy arrays to create the list of large
objects. We first create an array object_areas containing
the object areas, and an array object_labels containing the
object labels. The labels of the objects are also returned by
ski.measure.regionprops. We have already seen that we can
create boolean arrays using comparison operators. Here we can use
object_areas > min_area to produce an array that has the
same dimension as object_labels. It can then be used to
select the labels of objects whose area is greater than
min_area by indexing:
PYTHON
object_areas = np.array([objf["area"] for objf in object_features])
object_labels = np.array([objf["label"] for objf in object_features])
large_objects = object_labels[object_areas > min_area]
print("Found", len(large_objects), "objects!")The advantage of using NumPy arrays is that for loops
and if statements in Python can be slow, and in practice
the first approach may not be feasible if the image contains a large
number of objects. In that case, NumPy array functions turn out to be
very useful because they are much faster.
In this example, we can also use the np.count_nonzero
function that we have seen earlier together with the >
operator to count the objects whose area is above
min_area.
For all three alternatives, the output is the same and gives the expected count of 7 objects.
Using functions from NumPy and other Python packages
Functions from Python packages such as NumPy are often more efficient
and require less code to write. It is a good idea to browse the
reference pages of numpy and skimage to look
for an availabe function that can solve a given task.
Remove small objects (20 min)
We might also want to exclude (mask) the small objects when plotting the labeled image.
- Enhance the
connected_componentsfunction such that it automatically removes objects that are below a certain area that is passed to the function as an optional parameter.
To remove the small objects from the labeled image, we change the value of all pixels that belong to the small objects to the background label 0. One way to do this is to loop over all objects and set the pixels that match the label of the object to 0.
PYTHON
for object_id, objf in enumerate(object_features, start=1):
if objf["area"] < min_area:
labeled_image[labeled_image == objf["label"]] = 0Here NumPy functions can also be used to eliminate for
loops and if statements. Like above, we can create an array
of the small object labels with the comparison
object_areas < min_area. We can use another NumPy
function, np.isin, to set the pixels of all small objects
to 0. np.isin takes two arrays and returns a boolean array
with values True if the entry of the first array is found
in the second array, and False otherwise. This array can
then be used to index the labeled_image and set the entries
that belong to small objects to 0.
PYTHON
object_areas = np.array([objf["area"] for objf in object_features])
object_labels = np.array([objf["label"] for objf in object_features])
small_objects = object_labels[object_areas < min_area]
labeled_image[np.isin(labeled_image, small_objects)] = 0An even more elegant way to remove small objects from the image is to
leverage the ski.morphology module. It provides a function
ski.morphology.remove_small_objects that does exactly what
we are looking for. It can be applied to a binary image and returns a
mask in which all objects smaller than min_area are
excluded, i.e, their pixel values are set to False. We can
then apply ski.measure.label to the masked image:
PYTHON
object_mask = ski.morphology.remove_small_objects(binary_mask, min_size=min_area)
labeled_image, n = ski.measure.label(object_mask,
connectivity=connectivity, return_num=True)Using the scikit-image features, we can implement the
enhanced_connected_component as follows:
PYTHON
def enhanced_connected_components(filename, sigma=1.0, t=0.5, connectivity=2, min_area=0):
image = iio.imread(filename)
gray_image = ski.color.rgb2gray(image)
blurred_image = ski.filters.gaussian(gray_image, sigma=sigma)
binary_mask = blurred_image < t
object_mask = ski.morphology.remove_small_objects(binary_mask, min_size=min_area)
labeled_image, count = ski.measure.label(object_mask,
connectivity=connectivity, return_num=True)
return labeled_image, countWe can now call the function with a chosen min_area and
display the resulting labeled image:
PYTHON
labeled_image, count = enhanced_connected_components(filename="data/shapes-01.jpg", sigma=2.0, t=0.9,
connectivity=2, min_area=min_area)
colored_label_image = ski.color.label2rgb(labeled_image, bg_label=0)
fig, ax = plt.subplots()
ax.imshow(colored_label_image)
ax.set_axis_off();
print("Found", count, "objects in the image.")
OUTPUT
Found 7 objects in the image.Note that the small objects are “gone” and we obtain the correct number of 7 objects in the image.
Colour objects by area (optional, not included in timing)
Finally, we would like to display the image with the objects coloured according to the magnitude of their area. In practice, this can be used with other properties to give visual cues of the object properties.
We already know how to get the areas of the objects from the
regionprops. We just need to insert a zero area value for
the background (to colour it like a zero size object). The background is
also labeled 0 in the labeled_image, so we
insert the zero area value in front of the first element of
object_areas with np.insert. Then we can
create a colored_area_image where we assign each pixel
value the area by indexing the object_areas with the label
values in labeled_image.
PYTHON
object_areas = np.array([objf["area"] for objf in ski.measure.regionprops(labeled_image)])
# prepend zero to object_areas array for background pixels
object_areas = np.insert(0, obj=1, values=object_areas)
# create image where the pixels in each object are equal to that object's area
colored_area_image = object_areas[labeled_image]
fig, ax = plt.subplots()
im = ax.imshow(colored_area_image)
cbar = fig.colorbar(im, ax=ax, shrink=0.85)
cbar.ax.set_title("Area")
ax.set_axis_off();
You may have noticed that in the solution, we have used the
labeled_image to index the array object_areas.
This is an example of advanced
indexing in NumPy The result is an array of the same shape as the
labeled_image whose pixel values are selected from
object_areas according to the object label. Hence the
objects will be colored by area when the result is displayed. Note that
advanced indexing with an integer array works slightly different than
the indexing with a Boolean array that we have used for masking. While
Boolean array indexing returns only the entries corresponding to the
True values of the index, integer array indexing returns an
array with the same shape as the index. You can read more about advanced
indexing in the NumPy
documentation.
- We can use
ski.measure.labelto find and label connected objects in an image. - We can use
ski.measure.regionpropsto measure properties of labeled objects. - We can use
ski.morphology.remove_small_objectsto mask small objects and remove artifacts from an image. - We can display the labeled image to view the objects coloured by label.
Content from Measurements
Last updated on 2025-10-15 | Edit this page
Overview
Questions
- Question 1
Objectives
Use regionprops from scikit-image to extract object-level measurements such as area, count, centroid, and circularity
Measure the mean and standard deviation of an object’s intensity
Export processed data for further analysis as csv / txt files
Export result images as TIF files using BioIO
This is a new episode that didn’t exist in the original curriculum
It could take content both from
06-processing-segmentation.md and
07-postprocessing.md.
Content from Validation
Last updated on 2025-10-15 | Edit this page
Overview
Questions
- Question 1
Objectives
- Explain the concept of validation in image processing
- Critically assess the segmentation results using appropriate validation techniques
This is a new episode that didn’t exist in the original curriculum
Content from Capstone Challenge
Last updated on 2025-10-15 | Edit this page
Overview
Questions
- Question 1
Objectives
- Combine multiple processing steps into a reproducible workflow
- Apply an existing workflow to several images in a batch analysis
In this episode, we will provide a final challenge for you to attempt, based on all the skills you have acquired so far. This challenge will be related to the shape of objects in images (morphometrics).
Morphometrics: Bacteria Colony Counting
As mentioned in the workshop introduction, your morphometric challenge is to determine how many bacteria colonies are in each of these images:
The image files can be found at data/colonies-01.tif,
data/colonies-02.tif, and
data/colonies-03.tif.
Morphometrics for bacterial colonies
Write a Python program that uses scikit-image to count the number of bacteria colonies in each image, and for each, produce a new image that highlights the colonies. The image should look similar to this one:
Additionally, print out the number of colonies for each image.
Use what you have learnt about histograms, thresholding and connected component analysis. Try to put your code into a re-usable function, so that it can be applied conveniently to any image file.
First, let’s work through the process for one image:
PYTHON
import imageio.v3 as iio
import ipympl
import matplotlib.pyplot as plt
import numpy as np
import skimage as ski
%matplotlib widget
bacteria_image = iio.imread(uri="data/colonies-01.tif")
# display the image
fig, ax = plt.subplots()
ax.imshow(bacteria_image)Next, we need to threshold the image to create a mask that covers only the dark bacterial colonies. This is easier using a grayscale image, so we convert it here:
PYTHON
gray_bacteria = ski.color.rgb2gray(bacteria_image)
# display the gray image
fig, ax = plt.subplots()
ax.imshow(gray_bacteria, cmap="gray")
Next, we blur the image and create a histogram:
PYTHON
blurred_image = ski.filters.gaussian(gray_bacteria, sigma=1.0)
histogram, bin_edges = np.histogram(blurred_image, bins=256, range=(0.0, 1.0))
fig, ax = plt.subplots()
ax.plot(bin_edges[0:-1], histogram)
ax.set_title("Graylevel histogram")
ax.set_xlabel("gray value")
ax.set_ylabel("pixel count")
ax.set_xlim(0, 1.0)In this histogram, we see three peaks - the left one (i.e. the darkest pixels) is our colonies, the central peak is the yellow/brown culture medium in the dish, and the right one (i.e. the brightest pixels) is the white image background. Therefore, we choose a threshold that selects the small left peak:
This mask shows us where the colonies are in the image - but how can we count how many there are? This requires connected component analysis:
Finally, we create the summary image of the coloured colonies on top of the grayscale image:
PYTHON
# color each of the colonies a different color
colored_label_image = ski.color.label2rgb(labeled_image, bg_label=0)
# give our grayscale image rgb channels, so we can add the colored colonies
summary_image = ski.color.gray2rgb(gray_bacteria)
summary_image[mask] = colored_label_image[mask]
# plot overlay
fig, ax = plt.subplots()
ax.imshow(summary_image)Now that we’ve completed the task for one image, we need to repeat this for the remaining two images. This is a good point to collect the lines above into a re-usable function:
PYTHON
def count_colonies(image_filename):
bacteria_image = iio.imread(image_filename)
gray_bacteria = ski.color.rgb2gray(bacteria_image)
blurred_image = ski.filters.gaussian(gray_bacteria, sigma=1.0)
mask = blurred_image < 0.2
labeled_image, count = ski.measure.label(mask, return_num=True)
print(f"There are {count} colonies in {image_filename}")
colored_label_image = ski.color.label2rgb(labeled_image, bg_label=0)
summary_image = ski.color.gray2rgb(gray_bacteria)
summary_image[mask] = colored_label_image[mask]
fig, ax = plt.subplots()
ax.imshow(summary_image)Now we can do this analysis on all the images via a for loop:
PYTHON
for image_filename in ["data/colonies-01.tif", "data/colonies-02.tif", "data/colonies-03.tif"]:
count_colonies(image_filename=image_filename)
You’ll notice that for the images with more colonies, the results aren’t perfect. For example, some small colonies are missing, and there are likely some small black spots being labelled incorrectly as colonies. You could expand this solution to, for example, use an automatically determined threshold for each image, which may fit each better. Also, you could filter out colonies below a certain size (as we did in the Connected Component Analysis episode). You’ll also see that some touching colonies are merged into one big colony. This could be fixed with more complicated segmentation methods (outside of the scope of this lesson) like watershed.
Colony counting with minimum size and automated threshold (optional, not included in timing)
Modify your function from the previous exercise for colony counting to (i) exclude objects smaller than a specified size and (ii) use an automated thresholding approach, e.g. Otsu, to mask the colonies.
Here is a modified function with the requested features. Note when calculating the Otsu threshold we don’t include the very bright pixels outside the dish.
PYTHON
def count_colonies_enhanced(image_filename, sigma=1.0, min_colony_size=10, connectivity=2):
bacteria_image = iio.imread(image_filename)
gray_bacteria = ski.color.rgb2gray(bacteria_image)
blurred_image = ski.filters.gaussian(gray_bacteria, sigma=sigma)
# create mask excluding the very bright pixels outside the dish
# we dont want to include these when calculating the automated threshold
mask = blurred_image < 0.90
# calculate an automated threshold value within the dish using the Otsu method
t = ski.filters.threshold_otsu(blurred_image[mask])
# update mask to select pixels both within the dish and less than t
mask = np.logical_and(mask, blurred_image < t)
# remove objects smaller than specified area
mask = ski.morphology.remove_small_objects(mask, min_size=min_colony_size)
labeled_image, count = ski.measure.label(mask, return_num=True)
print(f"There are {count} colonies in {image_filename}")
colored_label_image = ski.color.label2rgb(labeled_image, bg_label=0)
summary_image = ski.color.gray2rgb(gray_bacteria)
summary_image[mask] = colored_label_image[mask]
fig, ax = plt.subplots()
ax.imshow(summary_image)- Using thresholding, connected component analysis and other tools we can automatically segment images of bacterial colonies.
- These methods are useful for many scientific problems, especially those involving morphometrics.