# What kind of numerical image features should I use to distinguish between the attached types of poses?

I've around $500$ binary (black and white) images, of which I attach $2$ samples. I'm working on a machine learning problem of classifying images like them. I'd like to know what image features I can use that'd be very distinct for distinct images of this type. I tried HOG features, but the difficulty is they're very high dimensional (even for a $64 \times 64$ image, the HOG feature dimension is $1764$), so the models will overfit. Hence the question. Thanks!

• I'm not that familiar with machine learning, but from what material I've read/seen and my intuition I'd say a convolutional neural network with subsampling layers should work well here. – orlp Jul 22 '17 at 5:47
• Hu / Zernike moments? Why don't you preprocess your images with skeletonization or other techniques? – Evil Jul 22 '17 at 17:02