Research On Salient Target Detection Method Based On Low Rank

Visual saliency is a fundamental research issue in neuroscience,psychology,and visual perception.As an important branch of visual saliency,salient object detection can locate and segment salient areas in a scene.It has been widely used in computer vision,such as object recognition,object segmentation,image compression and image retrieval.Salient object detection models are divided into two categories based on their prior knowledge: top-down and bottom-up.In the bottom-up object detection models,the low rank recovery model is a representative method.In recent years,the theory of low rank matrix recovery has been used into salient object detection and has achieved considerable results.In low rank recovery model,an image is represented as the combination of background regions and salient regions.However,when the similarity between foreground and background is high,the existing low rank-based detection models are difficult to separate them accurately,which leads to the degradation of decomposition performance.In order to solve this problem,two salient object detection algorithms based on low rank are proposed in this thesis.The main contents of this thesis are given as follows:(1)In order to separate the salient areas of the image from the background,a method via dictionary and weighted low rank recovery for salient object detection is proposed.Firstly,a low rank recovery model is constructed,and a dictionary is incorporated into the model,which better separates the low rank matrix representing background from the sparse matrix representing foreground.Secondly,saliency maps of color,location and boundary connectivity priors are obtained.In order to make better use of the prior information,adaptive coefficients are generated by their saliency values.Lastly,a high-level background prior is constructed,then the prior is merged into a weighted matrix and added to the low rank recovery model.Comparing with eleven state-of-the-art methods on six challenging databases,the experiment results show that our approach outperforms the state-of-the-art solutions.(2)According to low rank recovery theory,a method via trace representation and regularization for salient object detection is proposed to detect salient object more completely.Firstly,according to the Von Neumann Trace Inequality,a trace representation of matrix is used to obtain lower rank solution rather than the nuclear norm.Secondly,a Laplacian regularization is merged into model to reduce connection between sparse matrix and low rank matrix.Finally,the color,location and boundary connectivity priors are integrated into a weight matrix,which is incorporated into the matrix decomposition model.Comparing with fourteen state-of-the-art methods on six challenging databases,the experimental results based on Matlab show that our approach outperforms the state-of-the-art methods.