The Segmentation Of Complex Texture Image

Image segmentation is the important and the basic problem in image processing and computer vision. The segmentation result has strong influence on the following recognition and interpretation. Therefore, a lot of segmentation methods have been proposed. However a universal method doesn't exist until now.The significance, the function and the development of texture segmentation are discussed in this paper. The statistical and the space/frequency methods of feature extraction, segmentation methods based on fuzzy clustering neural networks are discussed also. The arithmetic that aims at the problems of texture segmentation is improved and the experiments have been done.The main work that have been done as follows:(1) Through the comparison of different fractal dimension, we find that the improved box-counted dimension and brownial fractal dimension fit texture segmentation. Better segmentation results are acquired when combined it with Laws energy in texture images and noisy images.(2) Directionality is one of the basic texture characters, In order to use this character, texture image is convoluted with Laws models. The parallel edges, vertical edges and diagonal edges are filtered and the correspond fractal dimensions are calculated. The fractal dimensions are classified and the image segmentation is completed. Experiment results have been given in this paper.(3) The estimated feature isn't accurate if window size is small when the segmentation makes use of SAR model. Otherwise the border of segmentation isn't exact because of the transition of feature at the border. Whereas wavelet transform have good local time-frequency property, scale-changed feature and good localization. The two methods are combined in this paper; the original image is segmented through coarse segmentation and refined segmentation. Good results are acquired.(4) The important problem in the RBF neural network is how to establish the structure. It can't learn the sample data well if the structure is too simple, otherwise the problem of excessive-fitting and the fall of generalization capacity is emergent.So a new RBF optimization algorithm is introduced by combining hierarchy genetic algorithm and multivariate linear regression. The complexity of the structure is felled and the learning speed is accelerated. This makes the segmentation result better.