| Image segmentation is the key step of image processing technique to imageclassification technique, so the study of image segmentation technique is of greatsignificance. The most commonly used image segmentation algorithms are based onthe regional level, such as region-based Gaussian mixture model, region-based MRFand IRGS algorithm. IRGS algorithm has higher segmentation accuracy because of itsprecise mathematical model, but it is based on intensity gradient which can notdescribe the texture characteristics and is more sensitive to noise which reduces theaccuracy of the segmentation algorithm. As texture gradient can describe the texturecharacteristics of image and has the ability in anti-noise performance. In order tosolve the problems of IRGS algorithm, we combine texture gradient with IRGSalgorithm to complete image segmentation. The main work is as follows:1. We present an initial region model based on texture gradient. Texture gradientcan describe the texture characteristics of image, and watershed algorithm isessentially a region growing segmentation algorithm. First we combine texturegradient with watershed algorithm to get initial regions of the image. Then weconstruct region adjacency graph to establish the context of relationship. Finallyregion model based on texture gradient is constructed. The experimental results showthat the watershed algorithm based on intensity gradient is very serious inover-segmentation, and watershed algorithm based on texture gradient and improvedtexture gradient has improved the situation of over-segmentation.2. We establish image segmentation algorithm on the basis of initial regionmodel based on texture gradient. We combine IRGS algorithms with initial regionmodel based on texture gradient to complete the image segmentation. To furtherenhance texture characteristics and the anti-noise performance of texture gradient,image should be filtered using non-local mean filtering before segmentation. Theexperiments show that IRGS algorithm based on improved texture gradient is betterthan it based on intensity gradient, and segmentation accuracy has been furtherimproved after the non-local mean filtering. |