Texture Image Feature Extraction And Clustering Algorithm

With the rapid development of information technology, texture analysis has already been focused in some related fields during the recent years.Texture ana-lysis is mainly applied to texture classification, texture segmentation, texture sy-nthesis, etc. In essence, texture clustering is equal to classify different texture pictures to unknown classes on basis of pixels of an image belonging to differ-ent regions. The textures belong to different regions is not only related to the known pixel gray but also have a close relationship with the distribution of gr-ay levels in a neighborhood. Texture clustering is considered as a compound of two questions: feature measurement and clustering. The methods of feature mea-surement include statistical methods,frequency domain, difference histogram,etc.This dissertation mainly focuses on the feature measurement and the clustering algorithms. It includes three parts: feature measurement of the texture, feature choice and the improvement of the clustering algorithms.First, we introduce some methods for texture measurement and have improved the methods for estimating the direction and roughness of the texture. In the end of this part, we measure 15 texture features of the Brodatz texture database based on gray-level co-occurrence matrix. Second, these features have been used K-L transform into a new domain and chose the main features for clustering. Third, we cluster the texture pictures of the Brodatz texture database. In this dissertation, we have improved two clustering algorithms.One is the improvement of K-means algorthim.The initial clustering state of K-means algorithm affects the final result which usually fall into local extremum. In this dissertation, we have improved the method for choosing the initial clustering state and have solved the problem well. The results show that the improved algorithm can increase the veracity.Another one is the improvement of SA (simulated annealing) algorithm. During the process of convergence, it is important to choose the parameters of the cooling schedule which has a great influence on the convergence time and convergence results. In this dissertation, we improved the SA algorithm for the aspects as follows: 1.Using SA algorithm for clustering and changing the problem of texture images clustering into the problem of optimization.2.Having improved the method for choosing the parameters of cooling schedule. We use changeable markov chain which is longer at higher temperature (having a large hunting zone) and shorter at lower temperature (having a small hunting zone).3.Adding the memory function to the SA algorithm and increasing the veracity of clustering. The data of the experiments show that the results of the improved SA algorithm are better than the traditional one. It cost less time than the traditional one and enhance the ability of clustering.