i'm implementing my first CNN for image classification in a field with very few research.

I'm aware i could extract feature and then try SVM, knn...

But i want to be sure CNN is not a viable solution.

I have already tryied different architecture, hyperparametrisation... Result have increase from 45% to 60%. I still have thing to try (data augmentation,dropout, transfer learning...)

Are there clue that can tell me CNN will not work fo my task, or should i try all method to be sure cnn is not viable?

For exemple in my case, i got 12 categorie. 8 are easy to classify but for the 4 other it's very hard, and all my modification significantly affect my accuracy for the 8 class, but just increase accruracy for the 4 other class a little

  • $\begingroup$ What is the resolution of your images? How many images do you have per class? $\endgroup$
    – Michael M
    Apr 29, 2019 at 17:29
  • $\begingroup$ original resolution is 200 000*200 000. This is binary image (pixel = 0 or 1 only). I reshape it in size 96*96 (i tried shape 196*196 but computation time way too long). this is very unbalance data set, i got like : class A = 800 images, class B = 300 images, class C D E = 150 images per classes, class F G H = 100, I J K L = 50. But for example i have 85% accuracy on class J and 40% for class C. $\endgroup$
    – akhetos
    Apr 29, 2019 at 17:33