Texture Segmentation Based On Anisotropic Diffusion

Texture segmentation is to divide the original texture image into nonoverlapping regions each of which has the same texture properties. Various kinds of texture segmentation techniques have been proposed and applied in many practical image processing fields, such as medical image processing, remote sensing, industrial automation, and machine vision.It is more difficult to segment texture image than other types of images due to the variety and complexity of texture structure. There is neither a commonly accepted definition of texture nor a satisfying mathematical model to describe and analysis texture accurately. General strategy of segmenting texture image is first to extract local features and then to cluster these features. Different from conventional texture segmentation methods, our methods are developed on the image filtering properties of partial differential equations. Combining nonlinear filtering and image segmentation methods, an unsupervised texture segmentation method is proposed which has a good performance and avoids the complexity and difficulty in texture modeling and analysis.The basic problem of developing the anisotropic diffusion for texture segmentation is how to use the diffusion to smooth edges within texture regions and meanwhile to reserve boundaries between the regions. In the first part of the dissertation, the diffusion properties of isotopic and anisotropic diffusion is analyzed and compared. And the essential problem of choosing an appropriate diffusioncoefficient is discussed.Based on the theory of total variation minimization, an anisotropic diffusion model is proposed to obtain a sub-image which is composed of homogenous regions and sharp borders from the original texture image. Combined with the active contour image segmentation method, this model can be applied to texture segmentation.But the variety of texture image that can be segmented by total variation based anisotropic diffusion model is limited by the mechanism that the segmentation is mainly based on the difference of the average gray level between texture regions. Much of other useful information about texture properties such as texture orientations and scales are lost during the diffusion filtering. In order to solve this problem, we extend our method to vector-valued image.A set of Gabor filters are used to extract texture features which can form a set of vector-valued image. The total-variation-based anisotropic diffusion model is extended to the vectorized diffusion for texture segmentation through those texture features. Experimental results of segmenting a variety of texture images demonstrate the effectiveness and applicability of this approach.