# Customizing REGIONPROPS With Your Own Measurements 12

Posted by **Steve Eddins**,

I saw a presentation last month that mentioned a user request to have the ability to customize `regionprops`. That is, a user wanted to be able to add their own measurement to `regionprops`.

Today, I'll show you how to do this yourself.

First, here's a brief recap on what `regionprops` does. The function computes measurements of image regions. Some of these measurements are based purely on a region's shape, while others incorporate pixel values within the regions. Here's an example using the coins.png sample image.

```
I = imread('coins.png');
imshow(I)
```

Let's convert this image to binary, using adaptive thresholding, filling holes, and removing small "noise" pixels.

bw = imbinarize(I,'adaptive'); bw = imfill(bw,'holes'); bw = bwareafilt(bw,[100 Inf]); imshow(bw)

You can count the "blobs" (object) yourself; there are 10 of them.

The simplest `regionprops` call, `regionprops(bw)` computes the `Area`, `Centroid`, and `BoundingBox` for each object.

s = regionprops(bw)

s = 10×1 struct array with fields: Area Centroid BoundingBox

But I don't think this is the best way to call `regionprops` anymore. You can now tell `regionprops` to return the results as a table.

```
t = regionprops('table',bw)
```

t = 10×3 table Area Centroid BoundingBox ____ ________________ ____________ 2635 37.133 106.85 [1x4 double] 1846 56.131 49.693 [1x4 double] 2672 96.199 146.05 [1x4 double] 1839 109.97 84.848 [1x4 double] 2744 120.37 208.73 [1x4 double] 2520 148.57 34.404 [1x4 double] 2589 174.83 120.01 [1x4 double] 2518 216.81 70.649 [1x4 double] 1857 236.03 173.36 [1x4 double] 1829 265.96 102.64 [1x4 double]

The table form is a lot more convenient for many tasks. For today's topic, one especially nice thing thing about tables is how easy it is to add your own variables to the table.

To illustrate, let's add a measurement that I've seen called `Roundness`. One definition for roundness is:

$R = \frac{4A}{\pi L^2}$

where $A$ is the object area and $L$ is the major axis length of the best-fit ellipse for the object. Here's how to compute roundness and add it directly to the measurements returned by `regionprops`.

First, note that both `Area` and `MajorAxisLength` are supported by `regionprops`, so let's start with those.

t = regionprops('table',bw,'Area','MajorAxisLength')

t = 10×2 table Area MajorAxisLength ____ _______________ 2635 60.08 1846 50.178 2672 59.792 1839 49.674 2744 60.374 2520 58.08 2589 58.676 2518 58.162 1857 49.77 1829 49.564

You access table variables using dot notation, like `t.Area`. Similarly, you can create a new table variable using dot notation and assignment, like `t.MyVariable = ...`. So adding `Roundness` to the table returned by `regionprops` is this simple.

t.Roundness = 4 * t.Area ./ (pi * t.MajorAxisLength.^2)

t = 10×3 table Area MajorAxisLength Roundness ____ _______________ _________ 2635 60.08 0.92945 1846 50.178 0.93352 2672 59.792 0.9516 1839 49.674 0.94893 2744 60.374 0.9585 2520 58.08 0.95118 2589 58.676 0.95745 2518 58.162 0.94772 1857 49.77 0.95453 1829 49.564 0.94798

Let's try this computation with an image containing objects that are not quite as round.

```
I2 = imread('rice.png');
imshow(I2)
```

bw2 = imbinarize(I2,'adaptive'); bw2 = imfill(bw2,'holes'); bw2 = bwareafilt(bw2,[100 Inf]); imshow(bw2)

t2 = regionprops('table',bw2,'Area','MajorAxisLength'); t2.Roundness = 4 * t2.Area ./ (pi * t2.MajorAxisLength.^2); head(t2)

ans = 8×3 table Area MajorAxisLength Roundness ____ _______________ _________ 138 23.594 0.31562 120 18.152 0.4637 169 28.123 0.27207 157 23.793 0.3531 284 43.757 0.18885 200 26.259 0.36929 141 21.647 0.38311 177 29.087 0.26636

I'm a big fan of the (relatively) new `histogram` function in MATLAB, so let's use it to compare our roundness numbers. I will follow the advice given in the `histogram` reference page for normalizing multiple histograms so that they can be more easily compared. I'll set the y-axis limits to [0 1], which is appropriate for probability normalization, and I'll set the x-axis limits to [0 1], which is the range for `Roundness`.

h1 = histogram(t.Roundness); hold on h2 = histogram(t2.Roundness); hold off h1.Normalization = 'probability'; h2.Normalization = 'probability'; h1.BinWidth = 0.02; h2.BinWidth = 0.02; xlim([0 1]); ylim([0 1]); title('Histogram of roundness (probability normalization)') legend('coins','rice')

There you have it. You can add your own object measurements to the output of `regionprops`. It's especially easy if you tell `regionprops` to return a table.

I'll leave you with this question, dear reader: Are there measurements you would like us to add to `regionprops`? I am aware of an enhancement request for the Feret diameter. What else would you like to see?

Get the MATLAB code

Published with MATLAB® R2017a

## 12 CommentsOldest to Newest

**1**of 12

Steve, good to share this. I often compute what I call circularity = perimeter.^2./(4*pi*area). I should also start using the table output – just too used to the old structure array (for now).

The new histogram() is nice. It could be a replacement for imhist() except for cases where you want to use standard bins 0-255 and 0-65535 instead of the automatically calculated bins that histogram uses. Do you think that imhist() and histogram() could be consolidated somehow into a single histogram function?

Also, MajorAxisLength is of limited usefulness because the definition (max length of an ellipse fitted to the shape) is not what people usually want, and that is the maximum Feret (caliper) diameter. And the min width would be the caliper diameter 90 degrees to the max caliper diameter. Do you think you could add this more useful measurement?

**2**of 12

IA—Thanks for your feedback. I will share it with the Image Processing Toolbox development team.

**3**of 12

IA—The following code using histogram produces a plot similar to imhist, except that it does not include the gray bar, and it does not do y-axis clipping:

I = imread('rice.png'); histogram(I,'BinLimits',[0 255],'BinMethod','integers') xlim([0 255])

**4**of 12

IA—I think of perimeter as a fairly noisy measurement, and it is squared in your circularity measurement. In your experience, is the circularity measure using the squared perimeter term useful?

**5**of 12

IA—One possible alternative to perimeter-based circularity is the “compactness shape factor,” which is based on area and second moments.

**6**of 12

IA—Regarding the pros and cons of different ways to compute orientation, Russ says: “There are a number of different [orientation] parameters that are used, including the orientation of the longest dimension in the feature (the line between the two points on the periphery that are farthest apart, also known as the maximum Feret’s diameter), and the orientation of the major axis of an ellipse fitted to the feature boundary. But just as the centroid is a more robust descriptor of the feature’s location than is the midpoint, the orientation defined by all the pixels in the image is often better than any of these because it is less influenced by the presence or absence of a single pixel around the periphery where accidents of acquisition or noise may make slight alterations in the boundary.” *The Image Processing Handbook*, 2nd ed, 1995, p. 490.

**7**of 12

hi!

I am a student and i am studying the property of osteocyte lacunae (called blob) in bovine femour.

i would like to find the volume of each blob.

is possible to calculate the volume of each lacunae with the function called regionprops? i used it in image 2D to obtain centroid and area of each image in 2D. if the answer is false, could you suggest another function that works for that problem?

thanks a lot

Luca

**8**of 12

Just to second what Image Analyst said

… “MajorAxisLength is of limited usefulness because the definition (max length of an ellipse fitted to the shape) is not what people usually want, and that is the maximum Feret (caliper) diameter. And the min width would be the caliper diameter 90 degrees to the max caliper diameter. Do you think you could add this more useful measurement?”

I am currently unhappy with the fitted ellipse axes – am now off to research Feret diameter to see if I can implement it for my project

Regards, Pete

**9**of 12

Pete—Thanks for your comments. I looked into Feret diameter computations after receiving Image Analyst’s comments, and I expect to blog about it soon, including some algorithm details and code.

**10**of 12

Maximum feret diameter is the same as the diameter of the minimum enclosing circle. There are some functions on the file exchange to calculate this, as it requires a search (in linear time).

**11**of 12

Hi Steve,

Thank very much for your fantastic blogs. Could you please suggest how could I label components serially for quantification of particle growth in a image stack? I can quantify the particle quite well in each image but need to label the particles constantly across multiple images ? with help of regionprops ?

Thanks very much.

Guang

**12**of 12

Guang—If you can use the Computer Vision System Toolbox, take a look at the documentation example for motion-based multiple object tracking.

## Recent Comments