I wanna apply image processing method to find these defect.

I have many image of same part.The aim is to find defected parts. For big defects, I can detect them. But for small defect such as shown in below images, I can't etect them. Do you have any idea that which should I implement and which variable I can consider to detect them?

[Walter Roberson]

Could we get the original images for these, without the red circles?

Answer 1

[Parsa]

Hi Omid

It's a kind of crack detection problem. According to your question, you have many images of the same parts with different types of cracks. So, employing machine learning techniques could be a good solution, in order to detect the cracks automatically. Based on inspections and also your knowledge about the types and form of the defects, you can specify tiny ROIs around each crack. These defects are different in size and orientation, so you will have a good dataset of ROIs with different shape of cracks, which are training dataset for learning procedure. Here is a nice open access article about applying machine learning to detect cracks :

Applied Sciences | Free Full-Text | Effective Crack Damage Detection Using Multilayer Sparse Feature Representation and Incremental Extreme Learning Machine | HTML (mdpi.com)

There is another solution comes to my mind. Specially for your images, it's obvious that if a crack exist, there will be a connection line between circles. It means that, at least, there is one non-circular feature in an image with the defect. So segmentation techniques could be employed. Please take a look at these links:

@doc:Measure properties of image regions - MATLAB regionprops (mathworks.com)

The Watershed Transform: Strategies for Image Segmentation - MATLAB & Simulink (mathworks.com)

Although the same question has been asked before as in the links below, for the cases, craks are on a flat surface without special patterns. But your images have a circular pattern. Any way, these links also could be helpful :

Matlab code for wall crack detection - (mathworks.com)

Image processing for crack detection and length estimation - (mathworks.com)

[Walter Roberson]

Because of perspective and differences in distances, the blobs are mostly non-circular anyhow.

[Parsa]

Thanks for your comment,you're right. So most the blobs are in ellipse form,in different lengths and orientations, but there is at least one different feature or area, contains the defect.

[Walter Roberson]

When I look at the second one, it is not clear to me that the cut goes all the way to join the two circles. I think the "area" of the cut might be less than the adjustment for perspective. It might be necessary to check the ferret distance or something similar, carefully checking for "hair"

Answer 2

[Image Analyst]

Have you tried just subtracting the image from a perfect one? If they're not perfectly aligned you might have to call imregister first before subtracting.