Based On Independent Component Analysis Of Texture Segmentation

Image segmentation is an important section of texture analysis. It's the basis of texture image understanding, the segmentation quality will directly decides the result of target recognition. Texture segmentation is one of widely researched texture analysis models, which has been used in many realms such as remote-sense image processing and medical image processmg in recent years. In this paper, Independent Component Analysis(ICA) was introduced and applied to do texture segmentation. The main content of this paper is:Firstly, ICA theory and FastICA algorithm were expatiated based on the kurtosis criterion. For texture images, the ICA texture basis images calculated with FastICA algorithm have multi-frequency and multi-orientation features and powerful to describe texture images, which is useful for texture feature extraction in image segmentation.Secondly, The performance of ICA basis images and gabor filter banks were compared. The ICA basis images are richer, they may be frequency and orientation selective like gabor filters, but they may also realize multiple pass-bands of multiple orientations in a single unit. Experiments show which have better performance than gabor filters in texture segmentation. In order to extend the spatial describing ability of ICA basis images, wavelet transform was combined with ICA, which make it be able to segment texture images with large textons.The third, Multi-scale Independent Component Analysis(MICA) basis images were created according to the frequency feature of texture image which extend the frequency describing ability of ICA basis images. The small size ICA basis images were used to analysis high-frequency components of an image and decide the transition position of different texture regions, the larger ones for low-frequency to distinct different regions on the contrast. Moreover, considering the conflict between texture segmentation quality and calculation time, the energy conversation MICA algorithm was designed which realize the multi-scale analysis to a texture image with small ICA basis images.The forth, considering the independent character between ICA basis images, a new idea of ICA basis images adaptive selection algorithm was offered. Based on the energy criterion, several primary ICA basis images were chosen to analysis original texture image with low dimension vectors, which save lots of computer time.