Image Saliency Detection By Multi-feature

The salient targets of an image refer to the objects or regions which are visually significant and attract more attention than other objects or regions from the visual system. Saliency detection aims to use computers to simulate human vision system and find the salient targets automatically. With the development of intelligent computer vision system, saliency detection of image has become a hot research topic in the field of computer vision and pattern recognition. Traditional saliency detection algorithms such as biology heuristic algorithm, spectral residual method and the detection method based on the graph have various defects in that they can not detect saliency exactly or completely. Thus they can not meet the demand of further applications.Generally, the salient object of a natural image has the property: certain feature of the salient object is relatively sparse compared to that of the whole image. Inspired by this observation, most recent works focus on detecting saliency by using sparse decomposition of certain features of an image. In particular, the feature vectors of all pixels are rearranged into a feature matrix, the feature matrix are decomposed into two a low-rank component and a sparse component. The low-rank component represents the highly correlated features of backgrounds while the sparse component represents the salient objects. However, one single feature can't describe the rich variety of natural scenes. In this thesis, we propose to use multi-feature and we present a new decomposition model. Specifically, we first segment the image into some uniform superpixels and extract multi-feature of the each superpixel. Then the multi-feature matrix is decomposed into a correlated part and a sparse part. Then we design a self-representation model for each superpixel by using all the superpixels feature to linearly represent each superpixel feature. In this paper, we use a nonconvex trace lasso to penalize the coefficients of the correlated part instead of the nuclear norm because the former can better capture the correlation structure of the feature data; on the other hand, we use a group sparsity to penalize the sparse component because it has the ability of feature selection. Finally, we use the group sparsity error of the multi-feature to define saliency of the image. Additionally, we present a method to remove some small isolated patchs, which is falsely detected due to the disturbance of background. This method removes the clutter interference of the significant results effectively. Extensive experimental results show that the detection of our method is better than some related methods not only in visual effect but also in the objective assessment.