Feature Extraction Of Texture Image And Subspace Segmentation Clustering

With the development of storage technology and popularization of network, image database is growing larger and larger. Thus, how to classify, store and transmit these massive image data has been an urgent problem to be solved now. As a special image form, texture has played a more and more significant role in image applications. Thereby, this paper mainly focuses on the research and analysis in terms of texture image feature extraction as well as subspace segmentation clustering.General process of texture image analysis and recognition usually include its feature extraction and analysis. For one thing, feature extraction can be showed by the main texture features drawn from high dimensional texture image so as to achieve dimensions reduction. For another, texture image analysis includes texture image recognition, classification and clustering. Generally, the process of clustering is based on feature extraction so that we can ensure the accuracy and low time complexity. On basis of the two objectives above, this paper aims to complete the targets as following: Firstly, this paper studies the existing method of feature extraction. According to the drawback of current texture descriptor that they don’t consider the correlation and redundancy in features, the higher character description dimension as well as the large calculation amount, this paper puts forward a rapid texture descriptor of 4 dimensions GBD. This kind of descriptor which is able to calculate the directional derivative of four directions has superiority in owning a better distinguishing ability over the existing CS-LBP, ICS-LBP, FCS-LBP, and ECS-LBP. Besides, and this kind of descriptor has a higher clustering accuracy in extracted texture and the feature dimensions can be reduced from the existing 16 dimensions to 4 dimensions. This method which bases on texture features can greatly shorten the time of clustering. Secondly, this paper researches on the texture image clustering according to the subspace segmentation, and puts forward a method which depends on particle swarm algorithm adaptively setting parameters of LSR. The method calculates analytic solutions via the robust segmentation of least-squares regression and solves the problem which needs set parameters artificially. After repeated experiments and analysis, this paper draws a conclusion that the optimum parameter is in connection with the noise in data. Value of parameter should be appropriately increased if the data is quite large. Hereby, this paper puts forward a fitness function and searches adaptively the parameters by the particle swarm algorithm (PSO). And on account of the experiment results, we can conclude that searching parameters with the fitness function can not only maintain a higher clustering accuracy, but also avoid the blindness of setting parameters artificially.