Texture Feature Extraction And Classification Based On Directional Filter Banks

Texture classification plays an important role in the area of image processing and computer vision. It has been of growing interest over the last years. Typically, texture classification can be divided into two sub-problems: feature extraction and classification. This paper mainly focuses on the feature extraction aspect of the texture classification problem. Recent years, the multi-channel approaches are of particular interest for texture analysis, because of their association with human visual system models of perception that involve scale and orientation. This method extracts information by passing the texture sample through a bank filters. The output of each filter provides information about different features of the texture. In this paper we apply a new directional filter bank (DFB) to the problem of texture classification. The directional filter bank fits perfectly in the multi-channel framework by providing excellent directional selectivity while remaining the property of maximally decimated.In the real world, the textures often have some type of distortion with respect to database content, such as skewing and rotation. One commonly found distortion is rotation. Hence the crux of the problem is to obtain feature sets that are rotation invariant (RI).In this paper, we have proposed a new method of directional feature extraction based rotation invariant texture classification, named GLCM/DFB joint method, which combines two methods on rotation invariant texture classification, GLCM, gray level co-occurrence matrix-based and DFB-based method. Firstly, we have implemented texture classification based on DFB. To get RI measure, we perform DFT-encoding on the resulting feature vector. Secondly, jointing with GLCM method, a new version-GLCM/DFB RI texture classification is proposed. Then, we choose the Bayes distance rule for classification, and build up our training set and testing set on Brodatz album. The experiment show that satisfied results can be achieved on both structural and stochastic texture.