Image Feature Extraction And Classification Based On Multiscale Geometric Analysis

In recent years, with the development of human vision system, Multiscale Geometric Analysis (MGA) has become an important method for people to understand the world and phenomena. It has been applied in many domains including image processing, and researchers have begun to develop the multiresolution texture analysis models. Multiscale analysis is the analysis on different scales (resolution). Researchers have proposed a number of innovations and improvements of texture analysis combined with the Multiscale analysis, such as wavelet, Contourlet, Brushlet etc. Extensive research has demonstrated that these approaches based on the multiresolution analysis could achieve reasonably good performance, so they have already been widely applied to image analysis and classification.In this dissertation, MGA based methods were discussed, and new image feature extraction methods have been proposed, which were applied to image classification.First of all, a new feature extraction method is presented called directional features in frequency domain, which are obtained by dividing the Fourier plane into sectors according to the frequency and direction. The proposed method has been used in the Brodatz texture album and SAR image classification experiments in order to verify the advantages of our method. The scheme of classifier is due to the fact that the correlation existing between each two feature channels of the same kind of texture can be seen as a distinctive feature between different textures. Compared with other feature correlation based methods, our method operates directly in the Fourier plane, which is more flexible in the capture of frequency and direction information. Moreover, the use of FFT transform is computationally more attractive.Secondly, with the consideration of GLCM (Gray Level Co-occurrence Matrix) and Multiresolution analysis features'limitations, a new algorithm named Multiresolu-tion Co-occurrence Matrix have been proposed, which combined the most widely used wavelet with GLCM. Multiresolution Co-occurrence Matrix (MCM) can well integrate the multiresolution property and spectrum information of the transform, and the structural information of GLCM. This paper also presents a new feature selection method through the statistical analysis on the properties of MCM and wavelet energy. The performance of the proposed MCM is measured through the classification test on Brodatz album. Experimental results demonstrate that MCM statistics outperforms other methods such as wavelet energy, GLCM. Lots of experiments on feature selection illustrate the effectiveness of the proposed method that can not only reduce the dimension of feature, but at the same time maintain the classification accuracy.At last, in order to better utilize the Multi-directional property of MCM, the ideas of Multiscale Geometric Analysis is combined with Multiresolution Co-occurrence Matrix(MCM), and a Nonsubsampled Contourlet (NSCT) based MCM feature is proposed. Furthermore, a feature selection algorithm is raised according to the new feature, in the hope of maintain the Multi-direction and Multiscale character, while reduced the dimension and computational complexity of the feature. Experimental results illustrated that this new feature not only has the advantages of MCM based on wavelet, but also can better explore the image's multi-directional characteristics and has better texture describing capacity.This work is supported by National Natural Science Foundation of China (No. 60505010, No.60702062, National 863 Program (2007AA12Z223) and Program for Changjiang Scholars and Innovative Research Team (IRT0645).