0
$\begingroup$

So I've got an interesting problem that I'm struggling with and I wanted to hear some ideas on possible solutions. The data is not public and I can't go into much detail.

The problem involves a binary image classification problem (Pass/Fail) for finding defects. Currently the image is split into 80 zones and each of these 80 zones have 7 features so each image produces a matrix of features (Haralick's Textures). I have class labels for each of the images and it is possible to provide a label for each of the zones on each image (whether a failable defect exists). The problem with doing this (besides being tedious and thereby lowering the possible training set size) is that it's hard to be consistent when grading small pieces of images.

What I'm currently playing with is doing and LDA/PCA transform on each of the zones (7-D feature to 1-D) and then a simple transform (i.e norm of the vector) to now go from the 80 zones to an overall image class. Hasn't been working to well. Images with gross defects get classified correctly but borderline images are muddled with pass images.

So as a kind of brainstorming session, I'd like to hear of some ideas I can play with. Thank you

$\endgroup$
0
$\begingroup$

Here is a very simple-minded idea that goes one step backwards from where you are (where you split each image into zones etc). It assumes that computation is cheap. Just use each pixel as a feature. Each image then either contains a defect or not. So you have one binary class label for each image and npixel features. There are many ML algorithms to choose from; random forest usually does well.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.