A Foreground Segmentation Method For Image Retrieval System Based On Color Feature

Image segmentation is a basic technology of modern computer vision task. It has a direct effect on the accuracy of retrieval results in content based image retrieval system. Therefore it is one of the most important parts in image processing. Image engineering according to the research content and the research methods, from low to high, can be divided into three levels:image processing, image analysis and understanding.Based on Mean Shift vector, in this paper we use the developed Mean Shift algorithm to cluster image features. With the convergent points, it can realize the image smooth and image segmentation. We select the proper kernel function to distinguish the contribution of sample points. According to the distance of feature and position between sample points, the values of sample points are specified different by kernel function. Calculating the Mean Shift vector of all pixels in the image, we can get an image convergence point map with which can the image be segmented into many sizes and shapes of the area.The image segmentation result of Mean Shift algorithm often exist the over-segmentation phenomena which leading to the crushing of foreground and background. In order to solve this problem, selecting the effective region merging method can reduce the over-segmentation phenomena. The traditional regional merging algorithm needs us to set the minimum threshold value. When the similarity of two areas is greater than the threshold value, areas merge. How to fit the threshold value to different applications are difficult. In this paper we present a area merging algorithm based on the maximum similarity of color feature. It realizes the separation of foreground and background.We introduce the user interaction information in this merging algorithm. With a few labels drawn by user for the foreground and background, all the non-label pixels can be automatically merged and labeled as either object or background. The key is maximizing the merging of background regions, which is equivalent to accurate reserve the foreground. The experimental results demonstrate that the algorithm is not only suitable for single target but also comfortable to multiple targets in image segmentation. Compared to the segmentation results of Graph Cut the algorithm in this paper has a better performance in some scenes.