The most common extracted features for image recognition

Question

I want to build a neural network, but because I have high resolution pictures, I rejected the idea of passing the entire image to the NN.

I was wondering what are the most common extracted features for use as input to the NN.

I know that features depends on the kind of images. But I guess I'm wondering what are the standard features. I'd also like to know about papers or books with information and/or examples about this topic.

Answer 1

Different categories of image features come to mind:

1. Color features such as color histograms which could for instance be in RGB or HSV space
2. Other histogram approaches, e.g. histogram of oriented gradients (HOG)
3. Texture features such as Tamura's or Haralick's
4. SIFT and SURF features are popular as well

Luckily libraries exist that provide access to many image features. Have a look WND-CHARM, they claim to support ~3000 different image descriptors.

Answer 2

One standard approach is to use a restricted Boltzmann machine to do the feature extraction, and then reconsider the RBM as a neural network and finish the training using back-propagation. See, for example,

G. E. Hinton, "To Recognize Shapes, First Learn to Generate images," Progress in brain research, vol. 165, pp. 535-547, 2007.

This is an example of automated feature extraction. It sounds like you are also interested in human-directed feature extraction, for which I am looking forward to other people's answers…