On Wavelet Features for Texture Discrimination

Texture features derived from wavelet transforms have recently been exploited for texture discrimination, image retrieval from a database, region classification for satellite images. Most works demonstrate that the error rate for texture classification is reduced as the number of texture features increases but seldom mentioned how to select good features derived from a specified wavelet transform. This report provides experiments to show that a few wavelet features might perform well if Whitney’s procedure is applied to select a suboptimal set of features. We test Daubechies four wavelet textures on three sets of database including (1) textures synthesized by Generalized Ising models (GIM), (2) textures synthesized by Gauss Markov random fields (GMRF), (3) natural textures scanned from Brodatz’s Album. A comparison with the features derived from Fourier transform, another filtering method, shows that both approaches can achieve perfect results if an appropriate set of features are used.