Texture Image Classification Based On Multiscale Geometric Analysis And SVM

With the increasingly using of the image database, it is more and more important to extract the usefulvisual information, which promotes the research of the efficient classification and retrieval of the images.Therefore, image classification has important value of practice and research. The research of theclassification of images mainly contains two aspects: the feature extraction and the classification. Texturehas rotation invariance and strong resistance for noise. It is an important feature for image classification.Multiscale geometric analysis is developed based on the wavelet analysis, and it is an efficient way toprocess the signal and the image. Because of its multi-directional and anisotropic, the feature of the imagecan be extracted more efficiently. Among the various transformation tools, Ridgelet transform, Curvelettransform, Contourlet transform, and Brushlet transform and so on, are representatives of the multiscalegeometric analysis. Brushlet is one effective image analysis tool to solve the angle-resolved problem.Support Vector Machines (SVM) is a global learning machine, and it is a way to realize the statisticallearning theory. SVM can achieve the principle of structural risk minimization. For resolving the linearinseparable problems in the original space, it maps the samples to the higher dimensional space, by usingthe kernal function. Based on the multiscale geometric analysis and SVM, this paper has done the researchas follows.1. This paper proposed one image classification algorithm based on complex feature from Brushlettransform. Since Brushlet is a complex function and is multidirectional, then we extracted energy and phasecharacteristics to describe the texture information. Experiments on Brodatz texture image database showthat this classification accuracy are better than GLCM and wavelet transform, the algorithm effectiveness on the texture image classification was verified.2. This paper proposed one multi-level pixel-based classification algorithm of images based on Gabortransform. For an image composed of different texture patterns, we could segment the image by classifyingevery pixel of the image. The feature of every pixel is characterized by its surrounding neighborhoods.The bigger size of the evaluation window could reflect the unity of the texture, and the smaller sizes reflectthe locality of the image. This algorithm could recognize the edge of the image, at the same time keep theunity of the image. Experiments on Brodatz texture image show that the algorithm is efficient to achievethe classification of images containing different texture pattern.