Image Saliency Detection Combining Foreground And Background Features Based On Manifold Ranking

In recent years, with the rapid development of computer and Internet technology, multimedia data is showing explosive growth. Thus, multimedia information storage, transmission and processing is becoming increasingly important, especially in image processing-related research and applications. Saliency detection is one of important preliminary works of image analysis and processing, and its goal is to detect the most salient and important part of an image. By detecting salient objects, the limited computing resources can be allocated to those salient regions of images with higher priority, which can greatly improve the efficiency of image analysis and processing. Saliency detection can be applied to numerous task s in computer vision fields, including content-aware image editing, object classification and recognition, image segmentation, image retrieval, etc. Therefore, image saliency detection has been widespread concerned in last few years.Saliency detection via graph-based manifold ranking(MR algorithm) is a very effective and representative algorithm in current studies. It considers the saliency detection problems as ranking tasks in a manifold structure. In MR algorithm, image boundary nodes are used as query nodes and the saliency map is obtained through two-stage manifold ranking. In most image scene, MR algorithm has good performance. However, since the MR algorithm being excessively dependent on background features of boundary nodes, when the internal features of salient regions are similar to the background features, or the features of each part of the salient regions have big difference, MR algorithm is unable to detect salient regions accurately. In order to solve the shortcomings of MR algorithm, two effective improved algorithm were proposed in the paper.Firstly, with regard to MR algorithm's overdependence on background feathers of the boundary nodes, a saliency detection algorithm combining foreground and background features based on manifold ranking(MRCFB algorithm) was proposed. The main idea of the improved algorithm was to take the global contrast foreground features in to account. For the question of how to combine the two saliency maps of foreground and background features, a query node selection method based on color and brightness was given in the paper. Regarding that in most cases MR algorithm performed well, query nodes generated by MR algorithm were still used in manifold ranking. When the salient regions detected by MR algorithm were quite different from regions computed via global contrast, both of foreground and background features were combined to obtain more accurate query node, and ultimately more accurate saliency map.Secondly, in the case of processing an image with complex backgro und, the disadvantages of MR algorithm become more obvious, and the MRCFB algorithm mentioned above do not have a better performance either, so a saliency detection algorithm based on the random walk for complex images(MRRW algorithm) was proposed. The algorithm was still based on the idea of combining foreground background features. Four saliency maps were obtained based on the foreground or background features via manifold ranking algorithm or random walk algorithm, finally these four saliency maps were combined to obtain the final saliency map of the image.Finally, MRCFB algorithm and MRRW algorithm had been tested respectively on two public dataset: Achanta(images with simple background) and DUT-OMRON(images with complex background), and then been analyzed in a variety of evaluation comparing with MR algorithm and other two saliency detection methods that have good performance around methods of recent few years. Experiments show that, 1) On DUT-OMRON dataset, MRCFB algorithm has almost same performance with MR algorithm. However, on Achanta dataset, MRCFB algorithm has obvious improvement in the precision rate, recall rate and F-Measure. 2) O n Achanta dataset, MRRW algorithm has almost same performance with MR algorithm. However, on DUT-OMRON dataset, MRRW algorithm's precision rate is improved by 8.5% and F-measure is improved by 5.7% compared to MR algorithm, at the same time, these two aspects are much better than other two methods. 3) Two algorithms in the paper have their own characteristics, MRCF algorithm is more suitable for images with general and simple background, MRRW algorithm is more applicable to images with complex background.