Texture classification by means of HMM modeling of AM-FM features

Abstract

This paper studies the classification problem of non-rotated and rotated textures digitized from the Phil Brodatz Album. The proposed texture analysis technique is based on AM-FM characterization followed by HMM modeling. The detection of AM-FM features was performed via a Gabor filter bank presented in a multiresolution way. To solve the problem of texture rotation, a technique was applied to correct the inherent orientation. In both cases, rotated and non-rotated textures, a low order feature vector was obtained from instantaneous AM-FM 2D maps. The proposed method was tested extensively and compared with some well-known approaches in the literature.