Some Researches On Image Filter Algorithm

There are four parts presented in this thesis. In the first part, the summarization of image filter is discussed. In the second part, the Statistical Learning Theory and Support Vector Machine are introduced. In the third part, the application of SVM is investigated in image filter in detail, and how to choose the learning sets is emphasis. In the fourth part, a novel image filter algorithm is proposed. And the third part and the fourth part are our contributions, and also the main content in this thesis.There is spatial correlation between the neighbored pixels, so there is the adjacent gray value. And we can build a 2-dimention surface for the image, then the pixels are just the samples. Based on this point, we introduce the Statistical Learning Theory or SLT and Support Vector Machine or SVM. In section three, we introduce SVM into image filtering. SLT is a small-sample statistics, where SVM is a new powerful learning method, and we can estimate the unknown gray value of pixel by the chosen samples. As a central part, how to choose the learning sets is investigated because it is very important for SVM to choose the learning sets. Then we do many simulations in different learning methods and get a good performance.The surface of the image disturbed by impulse noise displays many peaks or vales, so some gray values of pixels will be drastically different from gray values of neighbored pixels. According to the characteristic of impulse noise, a novel algorithm for detecting of impulse noise point from images based on directional derivatives are proposed in section four. Like in the median filter, a small template around a pixel is used to determine if the central pixel is corrupted. First the absolute value of 1-order difference between the central pixel and the rest pixels in the template is computed respectively. Then we compare the absolute value of 1-order difference with a given threshold T. At last we count the total number K which the absolute value of 1-order difference is bigger than the threshold T. If the value of the variable K isbigger than another given threshold r . the central pixel will be taken as a noise point. At last, lots of experiments are presented in order to compare our algorithm with these conventional filter algorithms, and show that our algorithm has good performance for removing impulse noise.