Color Image Segmentation Based On Multi-scale Rough Set And Weight-variable MRF

With the rapid development of information technology, the image processing has been applied into many distinctive fields such as medical diagnosis, meteorological monitoring, and military detection and so on, playing a more important role in modern life. As a basic technology of computer vision, image segmentation is one of the critical technologies in digital image processing. The effect of image segmentation directly affects the quality of image analysis. Generally speaking, there are some limitations on applications of modern image segmentation technology, such as the inaccuracy of the segmentation result, the time-wasting in segmentation process, and the excessive manual intervention. Therefore, the digital image segmentation is one of the problems in the field of computer vision technology, and it is as well as a heated issue for scholars both in China and abroad.The major content of this study is color image segmentation technology, aiming at abstracting the foreground of color images by an effective segmentation method. This paper introduces the basic theories and methods of color image segmentation, and proposes a revised segmentation method that is based on Markov random field. Compared to the traditional segmentation method, this new segmentation method has improved the accuracy of segmentation, while also significantly reducing the computation amount of segmentation model. This study utilizes the coarse segmentation and fine segmentation methods, and acquires over-segmented image by using multi-scale rough segmentation, reducing manual intervention. In the process of solving this segmentation model, the adaption of tabu search algorithm optimizes the result and reduces the amount of computation effectively.The color image segmentation of this study mainly contains two kinds of methods:coarse segmentation and fine segmentation. In the stage of coarse segmentation, we utilize the multi-scale rough method to realize the color image segmentation. First, under the conditions of a certain scale, we extract histogram and histon histogram from color images in three channels. We use those two histograms to construct a rough histogram for each channel and utilize the principle of maximum entropy to obtain optimal scale values for segmentation. Then, according to the rough histogram we obtained, we use the adaptive method to obtain the peaks and troughs of the rough histogram and split gray segmentation under each channel, realizing the color image segmentation. Based on this, we further utilize regional integration to reduce the number of areas to get more reasonable coarse image segmentation. In the stage of fine segmentation, we use the adaptive weights-variable Markov random field segmentation model to achieve segmentation. First, we extract color and texture features of color images, defined as HSV color values and rotation invariant LBP values, and we regard the result of coarse segmentation and the number of regions as the initial segmentation and number of categories of the segmentation model. Then, this study tries to solve the multi-featured, adaptive and weights-variable segmentation model, with the aim to find a label field that meets the maximum of a posteriori probability, using the characteristics of tabu search algorithm to optimize the solution process. Finally, based on area of the region and the region position, this study detects outliers and removes some error areas, thereby obtaining the final segmentation result.This thesis uses the experimental objects and standard segmentation results from Berkeley database to testify the effectiveness of this segmentation method. The experiments show that this method can achieve higher segmentation accuracy, effectively reducing the amount of computation as well as the inconvenience of manual intervention. In conclusion, this study has certain theoretical significance and practical value.