Co-Saliency Detection Based On Partially Absorbing Random Walks

With the continuous development of multimedia technology, the way of delivering information by image has been developed rapidly and digital image processing has become a problem. Salient object detection aims at finding interesting regions in images, since such regions contain important information and easily attract human attention. How, with the change and increase of the application scenarios, saliency detection for single image has been unable to meet the requirements of the object detections. Recently, there is an algorithm of detecting object in a set of images and this algorithm is defined as co-saliency detection. Co-saliency detection is used to discover the common salient object on the multiple images. Compared with detection for single image, co-saliency detection adds prominent object correspondence in other images as a new condition, which adds complexity to the problem.Considering that co-saliency detection stems directly from saliency detection for single image, this paper detects co-saliency from single image firstly. Different from previous methods, firstly, we optimize the similarity matrix of graph model by self-diffusion, the optimized similarity matrix can better capture the internal structure of graph. Then we apply partially absorbing random walks to saliency detection and combine the optimized similarity matrix, here we measure saliency by absorption probabilities. Finally, considering the differences between image pair and multiple images in co-saliency detection, we detect co-saliency with two different methods for image pair and multiple images. In multiple images, we select the nodes on boundary of image as background seeds, obtain initial saliency map of each image by partially absorbing random walks. Then we segment the initial saliency maps and exploit the statistical properties of co-salient object to select the co-salient seeds for each image. Finally we propagate saliency information throughout the graph. In image pair, we obtain foreground similarity probabilities which represents the inter saliency of multiple images by exploiting multi-scale segmentation, and then obtain background probability weight which represents intra-saliency of single image by absorption probabilities. Finally we combine two saliency cues efficiently by optimizing the objective function.We test effectiveness of the algorithm in this paper on several public benchmark datasets and compare with current popular methods, and the result demonstrates our methods achieve good performance.