I would suggest you to make more samples. Also, treat every image/sample as different sample (even they are same sample with different perspective/angle). You can use data augmentation and GANs to generate more samples, thereby, increasing your dataset. Keep this as a note, deep learning shines where you have alot of images/samples/data, therefore, better make more samples.

Since your subject is very niche, before trying GANs (because they are very experimental) I recommend you to understand your data very well. As I understood you try to identify cracks in an image which obtained from different perspectives. So I recommend you to start simple with a 2 layer neural networks and the aim of the classification must be labelled as 1 other is zero (assuming you do not try to differentiate the stages of cracks). To form a dataset from images you can simply try pytorch or keras, they have their own data loaders and you should prepare your data according to the data loader.

I hope this helps initially, if you give more concise examples I can try my best to help further.

Thank you so much for your replies,

I have tried building a small supervised problem using alex-Net in order to do that I used every single image as a different sample, Now here are the things getting very tricky.

Lets suppose I have a sample B, and it belongs to the class 0 that is bad sample with crack.

Now when I give the label to all 27 perspectives of this sample as class 0 then there are actually both perspectives, those where the crack is visible and also those where the crack is not visible. This is a random operartion so crack could be visible in any of the perspective.

So my network does not learns anything at the end, One thing could be done is that i manually go through all the bad samples and through away the ones in which the crack in not visible? any suggestions for this specific problem.

Thank you so much, Meltem Atay and Ajmal Shahbaz

Manav Madan Make it binary, when cracks not visible it is 0 when visible it is 1 and treat perspectives as different samples so sub 27 samples of B would give you 27 different data points which may be 0 or 1.

I do not know the implementation details but it should learn this way (if you have enough data and and all set correctly)

Try to make your work look like this sample: https://www.kaggle.com/sujit25/cats-vs-dogs-classification-with-cnn-keras

model starts at ln20 previous steps are preprocessing you can skip those if you want.