Research On Image Salient Object Detection

The human eye's visual attention mechanism could distinguish the salient object quickly and accurately in complex scenes,resulting from differences in eye irritation between different objects.Whereas,Salient object detection assigns different salient values to different things in the image on the basis of visual attention mechanism.As one of the most important part in image processing field,salient object detection attempts to detect the salient region in the image accurately,and it has a wide applications,such as image segmentation,image compress,image indexing and scene reconstruction et al.Methods of salient object detection can be summarized as top-down and bottom-up.Top-down methods rely on high-level information and the learning and training of neural network to detect salient object.The bottom-up approach rely on the comparison of features such as color,texture,and spatial distribution of the image to distinguish the background and salient object.However,methods based on contrast are difficult to form an effective saliency map mainly by highlighting a certain area or restraining a certain area.Thus,we propose the salient object method based on iterative contrast for RGB image.The propose method iterative contrast from the background and the foreground that they complement each other on visual attention to improve the effectiveness and accuracy.From the perspective of background views,the propose method according to the background priori and filtered the boundary through the pixel histogram to determine background label,then Initialize background iterative contrast with background label.Background iterative contrast can effectively suppress the region similar to the boundary and reduce the misidentification that the salient object appears on the boundary.From the foreground perspective,utilize the foreground label that determined by compactness calculation and threshold segmentation to initialize the foreground iteration contrast,and using the result of background iterative contrast as confidence corrected the result of foreground iterative contrast in every step to outstand object complete while suppress the background.With the development of distance measure sensors and the application of depth information in many industrial fields,salient object detection gradually extended from RGB images to RGB-D images.Many existing methods for salient detection of RGB-D image exploit the color feature of the RGB image and the depth feature of the depth image to calculate saliency respectively and fuse the results.The above method has achieved good results,but when the depth feature of the object and the background are at the same level,the saliency result obtained from the depth image will be useless or wrong,thus affecting the final test result.We believe that saliency computation alone using depth feature does not fully reflect the benefits of depth information and we propose salient object detection based on multiple perspectives fusion for RGB-D image.Multiple perspectives fusion framework includes fusion of color and depth,fusion of global and local and fusion of background and foreground.In this thesis,firstly,we fuse color and depth features at pixel level and super pixel level,and build a feature fusion graph model to avoid mistakes caused by individual color features or depth features.Then,the feature of color and depth fusion are used to guide the calculation of global and local compactness,and the results are fused to improve the accuracy of saliency detection.Finally,we according the visual difference between background and foreground to optimize the result of the previous step from background and foreground perspective successively to obtain the highlighted and uniform saliency map.Experiments on MSRA-1000,CSSD and ECSSD benchmarks show that the proposed salient object detection method for RGB image can effectively enhances the accuracy of detection and outperform other state-of-the-art methods in both PR curve and F-measure.Experiments on RGBD-1000 show that the proposed salient object detection method for RGB-D image outperform other state-of-the-art methods in both PR curve and F-measure.