3
$\begingroup$

i am thinking about using a CNN to classify certain images from industrial production, like scratches, stain, particles etc.

the problem is we don't have many images. We only get around 10 parts with certain defects to detect. I thought about using a database with different defects on it.

I use the example of scratches, there are around 100 images of scratches. Now my question:

Does it make sense to rotate the images in all 4 orientation, and to mirror them as well? So i would get 8 images (the original + 7 uniformly transformed images), i know the "value" of the rotated and mirrored images is not as high as from real images, but it should still help the CNN to abstract and find features.

What do you think about it?

Thanks in advance!

Dog_Fish

| improve this question | |
$\endgroup$
1
$\begingroup$

I am currently doing a project with similar industrial image processing task. I am detecting defects in a rectangular product image and from my experience, horizontal and vertical flipping of images has helped me a lot to augment my dataset and increase F1 score. Since I am not sure regarding your picture orientation I am unable to answer regarding other rotations. But if it has equal height and width then rotation of dataset makes sense. Try incorporating residual units into your CNN which has helped me massively since I had to use deeper networks to extract intricate features.

| improve this answer | |
$\endgroup$
  • $\begingroup$ I've just read about a network learning features for detecting skin cancer, in this project the author wrote: "To enlarge train data I tried image augmentation using flip. At first I tried to transform using flip, zoom and shift, but I thought it can twist training features much because medical images are very sensitive, so I tried to use flip only. However image transform didn't help this network and dataset, so I disabled it in this project." github.com/macgebi/udacity-dermatologist-ai/blob/master/… That's surprising, it is so intuitive that it helps $\endgroup$ – dog fish Dec 3 '18 at 7:08
  • $\begingroup$ yeah, that is why at end of the day, the augmentation technique depends on the project itself. One can only know if it would work by doing an augmentation and comparing the results. In traffic scenes prediction and food image processing personally I have had better results with flip augmentation $\endgroup$ – Sangathamilan Ravichandran Dec 3 '18 at 8:46

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.