TEXTURE ANALYSIS AND DISCRIMINATION IN ADDITIVE NOISE

Texture feature extraction and discrimination in an additive noise environment is considered. A new set of 28 texture features derived in the spatial frequency domain is presented. A successive addition and deletion feature selection scheme based on the Wilks criterion is used to obtain a subset of features which effectively discriminate among a set of sample textures. The selected features are further evaluated according to their ability to discriminate and classify samples of natural textures corrupted by additive noise at various signal to noise ratios (SNR). Samples from five texture classes were classified using only four features with an accuracy of at least 92% for all SNR greater than or equal to one. Performance for a SNR of 0.1 dropped to 70%. The four texture features extracted measure the dominant peak energy, power spectrum shape, and entropy. These measures are insensitive to additive noise and are effective for texture discrimination.