The Salienct Object Detection With Prior Integration

With the development of information technology and the popularity of digital products,image and video resources become more abundant,but the information redundant is still a challenge.Accordingly,extracting the key information from image and video resources and interpreting them effectively are actually more and more important.Inspired by the efficient processing mechanism of human visual system,the image saliency detection model has been emerged.The primary task of image saliency detection algorithm is to accurately extract the most salient regions which can attract a particular attention of human.Previous saliency approaches almost employed contrast prior or center prior individually.However,the salient object cannot always be highlighted uniformly,and the traditional center prior will become invalid when the salient object is far away from the image center.Thus,this thesis develops a novel bottom up salient object detection algorithm based on prior integration.In contrast to the most previous algorithms which processed low-level information of image directly,our proposed algorithm uses a two-stage strategy from coarse to fine to do saliency detection.In the first stage,we use a prior integration to generate coarse saliency map.Based on the existing problem of weak robust in traditional center prior,we apply the corner detection method to locate the center of the salient roughly.And according to the detected center,we expand the new center prior which makes the proposed algorithm more robust.In addition,considering that the salient object in image will not be cropped by the image boundaries,we define the image pixel saliency as the correlation of current pixel with different pixels on image boundaries to make the use of the boundary prior more fully.Moreover,the contrast prior and the new center prior as well as the boundary prior are combined to compute the coarse saliency map.To further refine the effectiveness of coarse saliency map,we propose an energy function based on prior integration in the second stage.The energy function is composed of the background energy,the data energy and the smooth energy terms which are based on the boundary prior,the coarse saliency map and smoothness prior respectively.The salient object in image can be highlighted accurately and uniformly after the treatment of the energy function.Finally,we use three public databases to validate the effectiveness of the proposed algorithm,including the validity of boundary prior,coarse saliency map and energy function.Experimental results demonstrate that the proposed model outperforms the most state-of-the-art methods.