Shape and data driven texture segmentation using local binary patterns

Official URL: http://risc01.sabanciuniv.edu/record=b1164872 (Table of Contents)

Image segmentation is a fundamental step in image analysis. Segmentation can be done by isolating homogeneous regions within an image or finding the boundaries between such regions. There are several cases that intensity values, color, mean or variance of image intensity distributions and edge information cannot play a discriminative role in image segmentation. While other features are not sufficient to discriminate regions, texture might be a good feature to handle the segmentation problem. Texture analysis and texture segmentation are still challenging problems; there is no method which can clearly identify and discriminate all kind of textures. Especially identifying nonumform textures and discriminating from other textures are still difficult problems in texture analysis. For this reason texture segmentation approaches may give unsatisfactory segmentation results with missing data or with corrupted boundaries of regions. Using prior information about the shape of the object can aid segmentation which can also obtain a solution to occlusion problems. In this thesis, we propose a shape and data driven texture segmentation method using local binary pattern (LBP). In particular, we train our LBP based texture filter with the texture which belongs to the region that we want to segment. We input the textured image into our filter to produce a "filtered image" which has been eluded from the structural properties of texture. Then by an energy functional, which combines the data term produced from the filtered image and shape prior term under a Bayesian framework, we evolve our level set based active contour for segmentation.