Rotation And Scaling Invariant For Texture Analysis

Content-based image retrieval (CBIR) is one of the most active hotspots in the fields of multimedia retrieval and artificial intelligence. It can directly analyze image content such as texture color and shape, etc., describe image features reasonably and make retrieval process more effective and more adaptive to human vision. Texture is one of most prominent character of image. Some important invariant features of texture, which are one of important technologies to improve the accuracy of image retrieval and classification, are applied to the image retrieval and classification. Texture invariant is an important study direction of texture analysis and texture classification, and plays an important role in many tasks, such as image segmentation, pattern recognition, shape analysis, texture synthesis and image compression.Texture technology are surveyed on the whole in this paper: the definitions of texture are summed up from different aspects, the study status of texture analysis and classification are summarized, and some general and new methods are introduced including kernel based image features extraction, statistics based feature extraction, model based texture analysis(fractal, random field etc.),structure based feature extraction and texture analysis, spatial-frequency domain based texture analysis(wavelet) etc. Based on the present study, two new methods are proposed and are applied to the texture classification successfully.(1) We noticed that the fractal feature in the nature can describe the structural complexity of the texture in some scale.However there are great limitation to comprehend the texture by the fractal dimension solely so that the accuracy of texture classification is not content to the application. The reason is that the fractal dimension is only the measure of the irregularity of surface which can not represent the curve sufficiently. So we combine the fractal model with statistic model to describe the texture curve. A rotation and scaling texture invariant is composed by the fractal dimension and two statistics. The texture invariant is applied to texture classification and by contrast with two other methods, this method exceeds in computation and accuracy.(2) After the Radon transform is performed on a texture, the features of the texture are kept in the statistics of the different orientations. And the principal line can be detected by the sensitiveness of the Radon transform to the brightness of the texture. A wavelet is applied to analyze the statistics, then a rotation and scaling texture invariant is composed. Texture classification is performed by the invariant successfully. Compared with other literatures, this method not only possesses of higher computing efficiency but also has better classification results.