Texture classification using wavelet frame decompositions

Multiscale approaches have provided researchers with a new avenue of approach concerning problems in texture analysis and segmentation. A new multiscale method is described one which extracts a feature vector that is based on a density of zero crossings of a wavelet frame decomposition for a particular texture. Each subband of the decomposition contains information for a particular scale and orientation. Thirty-four different textures were analyzed, with six different wavelet families (Haar, D4, D10, D20, DS8, and C6), and the classification results for different distance metrics are discussed. Also, the validity of this approach to the texture segmentation problem is addressed.