I have a dataset (300,000 samples) of grey scale images of dimensions 32x32. There are two classes and the samples are almost equally balanced. I am trying to achieve binary classification based on the variations in the image, especially on the boundary pixels, i.e., images which have mostly uniform intensity will be labeled as 0, while images with more variations will be labeled as 1.
I started with a small network but got an accuracy of 54% only. Thereafter, I started experimenting by adding more convolution layers, increasing filter size, reducing learning rate, changing optimizers etc. I used the modified AlexNet for CIFAR10 dataset as the CIFAR10 dataset also has 32x32 images only. The difference being in the number of image channels and the number of classes only. But, I could reached upto 64% accuracy only.
I understand that with two classes the minimum accuracy will be 50%. This implies that my network is not learning well enough. Now I need to know whether the problem is with my approach, because this may not be a simple classification problem. It may have some semantic information as well. I need to understand what training method or network is suitable for this kind of a problem. I'm using Tensorflow Tflearn.