I suspect you could just use the content loss approach from the style transfer examples. Take say the second last layer of a pre-trained convolutional neural n

I suspect you could just use the content loss approach from the style transfer examples. Take say the second last layer of a pre-trained convolutional neural network (e.g. VGG) and just use that to compute cosine similarities between images. Would have to pick a cutoff with some manual testing but I suspect that'd work plus can be done in a matter of hours using pre-trained weights in keras.