Research On Superpixel Image Segmentation Method Of Color Image

In the traditional method, the segmentation treat the pixel as basic processing unit, without taking into account the space and group relationships between pixels. This method will reduce processing efficiency of the algorithm when the size of target image is getting bigger and bigger. So, we propose the concept of superpixel. Through this method, we can effectively reduce the scale of the information that need to be processed and the complexity of the subsequent process. Currently, two advanced superpixel methods, SLIC and SEEDS, have some shortages and problems in the superpixel image processing. We propose some improvements to these two methods in this paper. The main research contents are as followed:(1) In the Simple Linear Iterative Clustering(SLIC) method, the number of superpixel needs to be set manually. So this number may not be very accurate. We can not ensure this value is the most appropriate one comparatively. This paper brings out a modified way to solve the problem above. To get the number of superpixel, we can use the color information of picture, scan some lines randomly and analysis the segmentation situation of similar color in pixel line. The simulation experiment shows that this method can help us to get the appropriate comparatively number of superpixel.(2) There's a precision problem in the boundary segmentation process in SEEDS superpixel color image segmentation method, so here comes a new one: Bilateral Filter SEEDS method. First of all, we treat the picture with Bilateral Filter method, this procedure can reduce the influence of texture and noise in boundary segmentation and filter the noise without missing the information of boundary. Now, the picture is smoother, and then, we treat the picture with superpixel color image segmentation, which reduces the erroneous in image segmentation, thus making the segmentation of boundary information more precise. Simulation experiment shows that the segmentation result of Bilateral Filter SEEDS method is better than the result of the SEEDS method. And also, the boundary recall rate and under-segmentation error rate are all having an advantage over the traditional method obviously.