Advice would be appreciated.
In a set of maybe 50000 images say 244 x 244 px (single channel). In each image I have a feature that occupies a small percentage say 5 to 10 % of the image. It can be seen by the human eye, though the bachground is textured and noisy. There are 10 different catagories of features across the data set and I need to classify them. I am sure this is a common problem say in medical imaging.
I need some advice on what would be the best way to way to attack this problem before I start collecting the data.
One way I thought might work would be to crop all the training samples so that the feature would occupy say 50% of the new training sample. Then once the model is trained use the cropped size to stride thru the test images hopefully classifying the feature when it lands on it.
However this seems a bit brute force (I suppose the accuracy depends on the length of the stride). I wondered if anybody had a better or more elegent suggestion/steer/diirection.