Currently using naive data augmentation based on image processing does not lead to a good augmented dataset. This is an issue when we work with wafermap images because only a small number of image manipulation techniques make sense for these kinds of images. Techniques such as horizontal and vertical flip are fine but gamma and contrast changes dont lead to accurate wafermap images.
So then we are left with a small number of image manipulation techniques that can be applied to a single wafermap image and this in turn leads to a much smaller augmented dataset. This is not enough to balance the classes.
So a much better option here would be to train a GAN based model to generate training samples for the imbalanced classes. At least its worth exploring.
Currently using naive data augmentation based on image processing does not lead to a good augmented dataset. This is an issue when we work with wafermap images because only a small number of image manipulation techniques make sense for these kinds of images. Techniques such as horizontal and vertical flip are fine but gamma and contrast changes dont lead to accurate wafermap images.
So then we are left with a small number of image manipulation techniques that can be applied to a single wafermap image and this in turn leads to a much smaller augmented dataset. This is not enough to balance the classes.
So a much better option here would be to train a GAN based model to generate training samples for the imbalanced classes. At least its worth exploring.