Texture Classification is the problem of distringuishing between textures, a classic problem in pattern recognition. Since many very sophisticated classifiers exist, the key challenge here is the development of effective features to extract from a given textured image.
Many approaches have been defined to extract features, such as Gabor filters or wavelets, however a great deal of recent work has focused on patch-based methods, whereby a texture is classified strictly based on a set of small patches of pixels extracted from a given textured image.
The difficulty is that too small a patch fails to be sensitive to non-local features of the texture, whereas a very large patch leads to enormous feature vectors and computational problems associated with very high-dimensional pattern recognition.
Recent work by Liu proposed using random linear functions of patches. The number of such random features needed turns out to be relatively modest, therefore it is suddenly feasible to do texture classification using large image patches, then with some number of random features. Astonishingly, such an approach outperforms the texture classification of finely-designed, state-of-the-art texture filters.
Clausi, D. A., “Towards a Novel Approach for Texture Segmentation of SAR Sea Ice Imagery”, 26th International Symposium on Remote Sensing of Environment and 18th Annual Symposium of the Canadian Remote Sensing Society, Vancouver, BC, Canada, pp. 257-261, 1996. Get it here.