Texture Classification Using VQ with Feature Extraction based on Transforms Motivated by the Human Visual System

The textured images are classified with a supervised segmentation algorithm (the classifier is trained first on texture samples). Texture features are extracted in the frequency domain and classified by a vector quantizer. Since feature vectors are computed and classified independently, for each pixel, a Kuwahara-like filter is used on the final classification to improve consistency. Feature vectors are computed by taking a 2D Fourier transform of a window centered on a pixel; the phase is discarded, while the magnitude of the transform is retained and mapped to polar coordinates. The polar mapping improves the precision of the classifier on the test problems. Principal component analysis is used to compact the feature vector and contain the complexity of the vector quantizer used to select a small number of representative signatures for each class. The training yields a small codebook associated with each texture image of the training set