Texture Classification Based On Wavelet Higher-order Statistics And Minimum Classification Error Probability

Texture image is a sort of frame, which is made up of some close conjoint elements. It has local invariable regulation and holistic disciplinarian. Texture classification is a basal question in researchful field of computer vision, which has important significance of research. Texture classification has many important applications in the fields of image segmentation, image retrieval, pattern recognition, machine vision, etc.Exceeding Gauss probability distributing is an essential characteristic of texture images; in addition, while selecting wavelet transform as theorem of texture transform, expression of wavelet of image takes on nonGaussia characteristic. Most of theorems of texture classification adopt two-order statistics of texture, for example of variance, to describe characteristic of textures. However, two-order statistics can not describe roundly nonGaussia characteristic. In order to solve this problem, we choose higher-order statistics, such as scenes and kurtosis, as character of texture. There are two ways of making use of higher-order statistics in describing texture. The first way is compact expression, which is to combine two-order statistics and higher-order statistics to describe texture. This way could heighten capability of expressing of distilled character, which can increase veracity of classification. Another one is overall way, which is to use histograms as character of describing texture. It utilizes probability distributing of filtered images to describe texture, which is in favor of expressing roundly texture. In this paper, the two ways were used.The performance of the designed classifier is closely related to the selection of the discriminant criterion. A desirable discriminant criterion should be consistent with the performance of the classification. That is, not only for the design samples, but also for the unknown samples. Moreover, the discriminant criterion should be analytically tractable and suitable for optimization over the parameters of classifier with a numerical search method. Several analytically more tractable loss functions are the perceptron criterion, the relaxation criterion and the minimum squared error (MSE) criterion. These discriminant criteria is that the decision rule is not embedded in the loss function in a functional form suitable for optimization. As a result, the discriminant criterion is not consistent with the objective of minimum classification error probability. To solve this problem, a new discriminant criterion, the MCE criterion, was proposed for the design of the discriminant function-based classifier. The MCE criterion incorporates the classification operation and approximates the classification error probability-based performance evaluation in a smooth functional form. Compared to the traditional discriminant criteria, the MCE criterion is more consistent with the objective of minimum classification error probability.Based above aspects, we research and impolder a new arithmetic of texture classification. The proposed method has been evaluated on the classification of 25 BrodTex texture, and more than 90% classification accuracy has been achieved.