Study Of Graph-based Color Image Segmentation Method

Image Segmentation is a process of partitioning an image into a certain number of disjoint regions, pixels in the same regions are similar with respect to these features such as intensity, color, texture and so on, while adjacent regions are significantly different in these features. As a low-level processing technology in computer vision, image segmentation plays an important role in image recognition, object detection and other computer vision technologies. For the slow speed of segmentation and easier to fuse when the color between background and object regions are similar, we produced a graph-based color image segmentation method based on images' color, edge contour and texture features. The algorithm of this paper contains several steps :Firstly, the image is preprocessed by Meanshift to reduce the segmentation's computational complexity, and a region adjacency graph is constructed with the over-segmentation regions produced by the preprocessing of the image.Then, the image is converted into Lab color space, which is used to measure the similarity between two adjacent regions; The edge contour is obtained by combining the local and global image information in order to get a complete object boundary in segmentation process; To reduce the over-merging resulted by the color similarity between the object and background region, the adjacent regions' similarity in texture features is measured based on a texture statistical model produced by Gabor filter and histogram.The Next step is to merge regions based on images' feature information and regions' spatial relations. As for the problem of the merging criterion and order of merging, the merging criterion is constructed by computing the similarity in color, edge contour and texture features, and the satisfactory regions are merging by searching the optimal merging-cost, which can preserve some global prosperity of an image during the process of region merging. Besides, the region merging is finished by refreshing regions' label.Finally,extensive experiment are performed on SED and ASD database in order to prove this algorithm's efficiency.Experiment results indicate that efficient segmentation result can be obtained by this algorithm even the color features between the background and object regions are similar.And this paper achieves better segmentation performance than other state-of-the-art segmentation algorithms.