In this demo app, I'll show some processing steps required for counting objects. Let's say you need to count food grains, like rice or beans, or maybe you need to detect nonstandard sizes for quality control.
The code for this application is written in C++, with the Qt5 library for building the GUI interface, and OpenCV 4 for image processing.
For this computer vision application, we will apply the following steps:
- load image
- convert to grayscale
- binarize the image (that is, make the foreground white and the background black)
- if necessary, clean small noise applying morphological operators like erosion/dilation
- count the objects with a connected-component labeling algorithm
There are two examples images in docs:
Let's start with a image with 50 rice grains.
Opening the image, you will find that there is a clear separation between the background and the foreground.
Converting to grayscale, we don't see much visual modification, but the 3 color channels are mixed into only one.
Now, we transform the grayscale image into a binary (only black and white) image, through a threshold operation. That is, if the intensity level of a pixel is less than a threshold value, it will be considered black, otherwise white. The default value in this application is 128 (from 0 to 255). Using the default threshold, we obtain a clear separation between the background and the foreground.
After that, we can apply a connected-component labeling algorithm that search all the image for blobs (white pixels that are connected somehow). This application draws a circle for each detected object. If the area of the object is more than 30% smaller or bigger than mean area, the circle is red, indicating lower confidence in the detection. If the area is between this 30% limits, the circle is green. The result for the rice image is the following.
So, we see that two clusters each with 2 grains are wrongly detected as one grain. We can improve the detection trying to dissolve the link between the grains with a morphological operators called erosion. It is like using a rubber around the borders of the object. We undo the counting (Ctrl+Z) and apply twice the erosion.
Applying again the connected-component labeling, we have only one missed count. You could try more erosion, combine erosion with dilation, or simply filter for a range of area values.
For another image, with 31 red beans, the process is almost the same, but we need to adjust the threshold value and invert background/foreground. Let's check it.
Opening the image, you will find that there are some shadows in the corners as well as glare on the beans.
Converting to grayscale, now we see a difference.
We can apply the threshold operation with the default value of 128.
We can see a lot of noise in the corners. So, it is better to adjust the threshold. Undoing the operation, we can set the threshold for 95 (in a future version, I'll add a histogram; there are also automatic thresholding algorithms).
The connected-component labeling algorithm assumes that the foregound is white and the background is black, so we need to invert the image.
And now, the result for counting the beans. It is a good result, with 3 errors from "glued" beans, 1 error from the glare (I was expecting more problems with it), and 1 in the upper left corner. Try to improve the results adjusting the threshold, and applying erosion and dilation!
There are more advanced techniques to apply, but I hope you've learned the basics with this tutorial.