Compact feature vector for natural texture classification

Gabor filter masks are used to extract succinct feature vectors from natural and synthetic textures. Textures from the Brodatz collection and real images, captured through a CCD camera, are used in the reported experiment. The results obtained represent a high accuracy of classification for those textures with pronounced orientation. A classification accuracy of up to 90% is obtained using a 50-feature vector. A lower classification accuracy is obtained with a 20-feature vector (about 70%), due to the fact that the orientation is not finely sampled so as to capture all the possibilities. These results compare well with results obtained using the same method on computer-generated artificial textures.