点击次数：

发表时间：2022-10-07

发表刊物：Jisuanji Xuebao/Chinese Journal of Computers

所属单位：电子信息与电气工程学部

卷号：43

期号：1

页面范围：16-28

ISSN号：0254-4164

摘要：With the explosive growth of image and video data, how to recognize and process image quickly and effectively has become a difficult problem in the field of digital image processing. In order to solve this problem, using the visual attention mechanism of human for reference, the image significance detection technology can accurately detect and extract the most important areas in the image, thus reducing the computational complexity of image processing and speeding up the process. Image saliency is widely used in image compression, coding, image edge and region enhancement, segmentation and extraction of saliency targets and other image processing problems. In recent years, the significance detection of image has attracted extensive attention and research, and a large number of significance detection methods have been produced, which are generally divided into two categories: bottom-up method and top-down method. However, the detection effect is greatly affected by the location of the target area, caused by the prior information such as the boundary priori or center prior in the field of image saliency detection, and the robustness is poor. In this paper, we propose a saliency detection method that fuses boundary connectivity with local contrast. Firstly, based on the local contrast of each pixel, the convex hull surrounding the foreground region is constructed and K-means clustering algorithm is used to enhance the foreground region and suppress the background region within the convex hull. Secondly, in order to obtain more accurate foreground probability, a superpixel map model was established and the accurate foreground in the convex bump was used as the absorption node of the random walk model to obtain the significance value of superpixel. Then, the significance value propagation within the cluster was used to calculate the foreground probability of superpixel. Thirdly, the edge connectivity is used to calculate the superpixel background probability. Finally, the foreground probability and background probability of superpixel are fused to obtain the final significance image. In addition, whether it is top-down or bottom-up, the existing classical algorithms have their advantages and disadvantages. A single method cannot consider all possible situations comprehensively, so it is necessary to study the significance detection fusion algorithm which is used to fuse different significance graphs to get closer to the truth graph results. In addition, we also proposed a fusion algorithm which is based on DS evidence theory. DS evidence theory is often used in the theory of uncertain events, showing its superiority in the fusion of uncertain events. To obtain better detection effect, we use DS evidence theory to fuse the results of several different significance detection algorithms at the pixel level. Experimental results with three public datasets, i. e., DUT-OMRON, ECSSD, and MSRA10K, show that saliency maps obtained by the proposed method are closer to the ground-truth maps. And the three evaluation metrics of precision-recall rate curve, F-measure value and average absolute error value are better than those obtained by twelve state-of-the-art methods. At the same time, it can be seen from the experiment that the results obtained by combining multiple algorithms with DS evidence theory are better than the results obtained by the detection of each algorithm. © 2020, Science Press. All right reserved.

备注：新增回溯数据