Study On Texture Image Segmentation Based On Clustering Analysis

Image segmentation is an important means of extracting target information, is thekey step of image processing. As a research hotspot, researchers have presented kindsof theories what can be used in image segmentation. Thereinto, the imagesegmentation method based on clustering analysis has been widely studied, obtained abetter segmentation result.Clustering analysis is a method of data mining. The clustering, is to classify dataunder some features, make feature similarity as large as possible in class, on thecontrary, as small as possible between class. It is an efficient unsupervised method.The advantage of clustering lies in that it doesn’t need a priori knowledge, it onlyneeds to set an objective function and obtain the clustering result through iterativeoperation. At present, the clustering algorithms proposed by researchers have beenwidely used because of their high efficiency and convenience. For some specificproblems, these theories still need to be improved.This paper studies on segmentation of texture image. Because wavelet transformprovides a comprehensive analysis method that integrates multi-scale, spectrum,structure and statistical basis for texture analysis. The method has better spatial andfrequency decomposition characteristics, it can make a multi-scale decomposition onimage through dilation and translation operations, is a multi-scale and multi-channelanalysis tool for image in excellent performance. The method provides an accurateand uniform frame for extracting and expressing features of texture image in differentscales. So this paper applies wavelet transform to obtain image features, then makes atexture image segmentation combined with clustering algorithm. The maininnovations are in the following three aspects:(1) Build a multi-scale pyramid structure of texture image based on wavelettransform, extract high-frequency and low-frequency feature information in each level,then make a image segmentation based on clustering from the coarsest scale, transferthe segmentation result to a finer scale combined with Relief algorithm, segment theimage level by level till it can obtain a segmentation result in the finest scale. (2) The mean shift algorithm has been deeply studied in this paper, thecomplexity has been discussed. The sensitivity of bandwidth is improved and theproblem of clustering number is solved. It realizes a finer segmentation of textureedges combined with multi-scale feature structure of texture image in wavelettransform.(3) Combined with wavelet transform and constraint mean shift, extractingresidential areas and urban tree canopy in high resolution remote sensing image.