Multiscale Texture Image Segmentation Based On Wavelet Transform

Texture image segmentation plays an important role in image processing and machine vision. It connects low-level visual applications with high-level visual applications. It is widely used in medical image processing and remote sensing image analysis today. Texture image segmentation refers to the process of partitioning a digital image into multiple segments according to the properties of different texture. The pixels in a region are similar with respect to some texture characteristic, adjacent regions are significantly different with respect to the texture characteristic. Texture feature extraction and classifier are crucial in texture image segmentation. In this paper, an investigation was made into statistical texture model based on wavelet transform, statistical texture feature extraction and classification and multiscale texture image segmentation.Firstly, the background and significance, the nowadays study situation, some hot issues and difficult problems of texture image segmentation were illuminated. And the process of texture segmentation was explained briefly.Secondly, the conception and definition of texture, some common texture features and the ways of feature extraction was introduced, such as gray level cooccurrence matrix, Tamura feature and texture features based on wavelet transform. After that some common classifiers were illuminated.Thirdly, textural statistical feature based on wavelet transform was introduced in detail. A comparison of different statistical features was made to find the better way in extracting better statistical feature. Human vision was sensitive to both spatial and frequency domain, texture analysis only based on spatial or frequency domain could not get a satisfying result. Texture features based on jointed spatial and frequency domain could characterize different textures effectively. Discrete wavelet transform could provide texture analysis in jointed spatial and frequency domain at multiscale. Although discrete wavelet transform had good spatial and frequency properties, it had its own drawbacks, such as shift variance and lack of directional selectivity. Texture feature of the wavelet transform domain changed a lot when the image shifted slightly, this resulted in blur edge and low precision. In this paper, more accurate texture feature was extracted by Gaussian mixture model of complex wavelet transform domain. And it improved the segmentation result by the experiment. Finally, research was made on multiscale texture image segmentation. The size of classification window is crucial to traditional segmentation. A large window usually enhances the classification reliability but simultaneously risks having pixels of different classes inside the window. Thus a large window produces accurate segmentations in large, homogeneous regions but poor segmentations along the boundaries between regions. A small window reduces the possibility of having multiple classes in the window but sacrifices classification reliability due to the paucity of statistical information. So a small window is more appropriate near the boundaries between regions. Multiscale texture image segmentation makes use of the segment of both small and large windows and gets a better segmentation. In addition, a Post-Processing algorithm was proposed to improve the performance of the multiscale fusion algorithm.The process of the multiscale texture image segmentation algorithm proposed in this paper is as follow: firstly statistical texture feature and the raw segment can be gained by maximum likelihood classifier at multisacle due to the multiresolution of wavelet transform. Then interscale fusion algorithm was done based on the raw segments from coarse scale to fine until pixel level to the multiscale-fusion segment. At last, the Post-Processing algorithm was processed on the multiscale-fusion segmentation to get the final segmentation. The experiment shows that the algorithm proposed in this paper is more robust and the segmentation results are improved.