Saliency Detection Based On Visual Attention Mechanism

With the popularity and rapid development of computer networks,a large number of digital images are produced every day.The definition of image saliency detection is extracting most attracted attention from images by using image processing methods.Visual attention mechanism refers to the human ability to select and process the area in visual scenes.Efective management and the establishment of efficient image retrieval method gradually become the focus of the research.If salient areas can be extracted from images and videos accurately and computer resources can be allocated to these salient areas,the efficiency of image analysis methods will be improved greatly.This paper studies feature extraction and saliency calculation method according to the basic principle of human visual attention mechanism.The main research works are listed as follows:(1)In order to detect salient region accurately in the image,an image saliency detection based on boundary and center prior in natural scenes is proposed according to the basic principle of human visual attention mechanism,.The original image is first segmented into superpixels using SLIC(Simple Linear Iterative Clustering)segmentation algorithm so that the detection result can keep the shape of the objects in the image.Then according to the theory of background prior,the background and salient object is roughly separated.Finally,the final saliency map which further highlight salient object is generated by regarding the centroid of the background prior saliency map as the center of salient object,which overcomes the problem that the traditional center prior fails to detect the target which deviated from the center of the image.Simulation Experiments demonstrate that this method can highlight saliency object uniformly and suppress the background in natural scenes effectively.(2)A multi-scale image saliency detection fusing context information is proposed to overcome the problem of false detection using single scale sparse reconstruction.The original image is first segmented into superpixels using multi-scale SLIC(Simple Linear Iterative Clustering)segmentation algorithm,and a background template is established using a sparse representation algorithm for sparse reconstruction.Then the image context information is used to calculate the salient value,which helps to smooth similarity between image patches of sparse reconstruction error and to eliminate the false detection that is caused when foreground image blocks are included in the background template.A weighted fusion strategy is designed to complete the multi-scale significant fusion.Finally,the position prior is added to make context saliency detection results more accurate and get the final saliency map.Experimental results show that the proposed algorithm can highlight saliency object uniformly and suppress the background in natural scenes effectively on the public standard salient object detection database.(3)In order to reflect salient regions in video sequence accrrately,a spatiotemporal saliency detection integrated with motion characteristic is proposed.First,use SLIC segmentation algorithm to segment each frame into superpixels.Then,compute spatial saliency by using manifold ranking method,and use optical flow vector regional construct to compute temporal saliency.Then,use an adaptive fusion strategy to merge the temporal saliency map and the spatial saliency map into final spatiotemporal saliency map.Finally,adaptive threshold method is used to extract salient regions from saliency map.Simulation experiment result demonstrates that this method is able to extract salient target with clear outline in dynamic scenes.