Recommendation for autoencoder for photo feature extraction

Question

Given 6000 40 X 40 photo patches taken out of 50 x-ray scans, what can be best way to extract useful features out of this patches? I need the method to:

1. not be too computationally costly

2. the latent space has to be a vector

I came across multiple kinds of autoencoders: fully connected, convolutional, fully connected variational, convolutional variational, denoising, deep belief networks and so on. I do not have the time to try them all out. So what would you recommend as most likely to work? and do you know any good implementation of it?

Answer 1

If it must be an autoencoder, I would recommend trying a adversarial autoencoder as suggested here: https://arxiv.org/pdf/1511.05644.pdf Such autoencoders decompose the images into useful features and create a smooth representation space.

When considering the architecture of the encoder, decoder and discriminator, it would probably be best to use convolutional neural networks. They contain less parameters than fully connected networks and thus this should be computationally feasible for 40x40 images. There are of course many types of convolutional networks, make sure that you have some type of regularization mechanism (such as batchnorm/dropout). You can find many example of suitable convolutional networks for images, eg implemented already on github.