Salient Object Detection Via Multi-feature Fusion

Salient object detection aims to automatically identify the most attractive object in a scene.As a foundation of various applications of computer vision and multimedia,salient object detection has attracted many research interests in recent years,and been extensively explored.However,the performances of existing salient object detection methods for RGB images do not rival the Human Visual System.An important reason is that human eyes capture depth information and motion information,besides color information.Various features complement each other to give more comprehensive descriptions with regard to visual saliency,which provide the opportunity to improve the effectiveness of salient object detection.In this paper,we propose novel salient object detection methods for RGB-D images and videos via multi-feature fusion.It is able to cope with the challenges faced by salient object detection based on multiple features,including the conflicts among saliency estimations of different features,as well as the inaccuracy of depth and motion information.Firstly,we utilize different features to detect the saliency object.Then,we fuse them together to generate a saliency map with high precision and low recall,so that common salient regions are found.Finally,the complete salient object is detected by saliency propagation based on the similarity measure of integrated features.Specifically,the proposed RGB-D salient object detection method segments the input into super-pixels by an extended SLIC algorithm.After that,two saliency maps are individually estimated by color feature and depth feature,and then fused together with refinement to produce the initial saliency map.In a graph model composed of super-pixels,finally,saliency values are propagated to generate the final saliency map.Similarly,the proposed video salient object detection method matches the super-pixels in consecutive video frames by optical flow,and construct a saliency propagation network,which is then initialized by the fusion of color-based and motion-based saliency maps.The final saliency map is generated by iteraively propagating the saliency values in the saliency propagation network.Extensive experiments on public datasets show that the proposed method,which is composed of effective procedures,outperforms the state-of-the-art methods.