I want to classify watermarked and non-watermarked images. Watermarking can be anything like a logo, letters. I have started with keras(with Theano as backend) and CNN. As the first step, I am using the same watermark and same images with or without watermark. I have 72 images in both categories. Epoch is 30. But I am getting an accuracy around 50%. How can I improve my accuracy?
I do not know how you did your network, but with keras but with 72 images this will be difficult to build a network. The watermark should generatse one pattern I guess it text like and therefore you should have a regularity in the pattern specially if you use preprocessing on the image to come back and with.
After this you could try with simpler methods to build classification, but with 72 images your training set will be really small you will have to do some reduction.
Hope this help