Saliency Detection Via Objectness Foreground And Background Prior

As an important branch of the computer vision,saliency detection can quickly obtain the most attractive attention of the region in an image,which is often used in many fields,such as image annotation and retrieval,object recognition,automatic image cutting and so on.A method of image saliency detection via objectness foreground and background prior is proposed in this paper to solve the low accuracy which is caused by insufficient density of salient region and incompleteness of significant object contour.Firstly,the image is segmented by super pixels,and the color space distribution of the super pixels is calculated as the low-level feature saliency map.Then,a number of candidate target windows are obtained by using a objectness proposal method,and the object information is obtained.The search area is set up by the windows,and the target windows are optimized by combining the normalized low-level feature saliency map.Further,we use multi-window features to predict the foreground objects of super pixels and obtain foreground saliency map.Secondly,the background template is set up by the image boundary super pixels,and each template subset is clustered by K-means algorithm.The updated template features are used as the sparse dictionary to calculate the sparse reconstruction error to obtain the background prior map.Finally,the final salient map is obtain by fusing the two salient maps with weight.This paper verifies the performance of our method on two public data sets of MSRA10 k and PASCAL VOC2007,and our method is compared with 11 methods.The experimental result shows that our method effectively enhances the internal density of the salient region,and the salient region has clear contour,and the salient object is more complement.As the result,the accuracy of the saliency detection is improved.The performance of our method is improved significantly,while ensuring the efficiency of operation.Especially,when the image has complex background,the method proposed has obvious advantages.