Texture Image Segmentation Based On Second Generation Bandelet Transform

Now the methods based on wavelet multi-scale analysis prevail in image segmentation. The basic idea of segmentation making use of wavelet multi-resolution is to decompose the image into low and high frequencies in different resolutions and extract features to segmentalize the image. However, because wavelet transform has the deficiencies of poor directionality and processing only point singularity, wavelet can't make good use of geometrical characteristics existing in the image, which promotes the emergence of more excellent multi-scale analysis methods.Bandelet is a new multi-scale geometrical analysis tool for image representation. This paper studies on second generation bandelet on the basis of first generation bandelet. It is well known that discrete wavelet transform offers high-frequency sub bands on three directions in each scale, but these three directions can't represent all the direction information contained in the image. The deficiency of poor directionality in wavelet transform goes against for better image representation. Aiming at this shortcoming, bandelet considers the geometrical directions of the image. The new method is an image representation technique based on edges of the image, whose remarkable aspect is the introduction of the geometrical flow. The main features of bandelet are that bandelet can make good use of the geometrically regular directions and adaptively construct the best basis function. If the geometrical regularity is known in advance to be made full use of, it can help a lot in processing the image. This paper analyzes the similarities and differences between two successive generations and proves that second generation bandelet is an excellent tool of making good use of geometrical directions reducing complexity in calculating geometrical flow and wrapped process for image representation. A new method of texture image segmentation adopting the quad-tree decomposition in second generation bandelet transform to obtain the best geometrically regular directions is proposed. This paper uses k-means clustering method and FCM clustering method separately to carry out the experiments. The experimental results show that the new method utilizes the geometrical characteristics of the image successfully in image segmentation.