Broken screen detection help!

I was solving a problem recently for an insurance company in which I have to build a model to classify a Phone screen into broken or non broken followed by tempered vs screen glass classification. I have tried lot of fine tuning with different architectures but it seems that the models are Biased towards one class I spite of the training data containing equal no of classes. As both the classes are very identical.

Can you please help me out?
I’m attaching training data link

Thank you so very much

(not sure)
Actually I guess the problem is that a scratch or broken lines in a screen which is black will hardly be visible…
(Have seen some images of both classes)

We need to figure something else out

I tried with fastai library… doesn’t seem to get good results in this case

deadline is nearing, can you help me out ?

I am not that comfortable with CNN’S but found this paper not sure whether it will work or not

Imo you might want to try basic Image Processing Techniques rather than DL here.