| In the past decades, people have paid much attention to the problem of texture analysis because of its fundamental importance and wide applications to computer vision, image processing and pattern recognition. Recently, many approaches motivated by human visual system such as multichannel decomposition are widely applied to texture analysis. Obviously, a good model for texture analysis should be able to utilize the information and interaction among the channels. In this dissertation we have combined two techniques, multichannel decomposition and hidden Markov model (HMM), for texture analysis. The HMM is used here to utilize the information of interaction among the various channels created by multichannel decomposition. This combined technique is used to solve two outstanding problems of texture analysis, robust texture identification and automatic unsupervised texture segmentation.; For the first problem, we have proposed a rotation and gray scale transform invariant texture recognition scheme. First, the textured image is decomposed into subbands via wavelet decomposition. The gray scale transform invariant features are subsequently extracted from each subband. The features from the subbands are arranged in a sequence which is then modeled by an HMM. One HMM is designed for each class of textures. During classification, each texture is scored against all the models; and is identified with the model yielding the highest score. For the latter problem, we have described a completely automatic texture segmentation scheme that needs no a priori information regarding the number of textures present in the image. Once again, each texture is modeled as one HMM. Thus, if there are M different textures present in an image, there are M distinct HMM's to be found and trained. Consequently, the unsupervised texture segmentation problem becomes an HMM based problem, where the appropriate number of HMM's, the associated model parameters and the discrimination among the HMM's are the foci of our scheme. The true number of textures (HMM's) can be found by iteratively grouping the homogeneous HMM's. Finally, extensive experimental results are reported for both the schemes. These results demonstrate the success of the proposed schemes. |
