Sparse Gradient Based Structured Matrix Decomposition For Salient Object Detection

Simulating to the selective mechanism of human vision,salient object detection technology automatically detects the most attractive and eye-catching targets or regions in images,which has important application value and theoretical significance in the field of computer vision.With the rapid development of this technology,salient object detection has been widely used in image segmentation,image compression,object discovery,image re-targeting,object recognition,etc.In the past few years,the theory of low-rank matrix restoration has been widely concerned by researchers in image research field.Many salient object detection models based on low-rank matrix restoration have been proposed successively.These methods are mainly based on a common assumption that the background part in one scene usually contains redundant information,while the salient region is often sparse.Therefore,the feature matrix of an image can be regarded as a combination of low-rank matrix and sparse matrix,corresponding to the background and salient region of the image,respectively.Based on this theory,a non-convex low-rank structured matrix decomposition model and a sparse gradient based structured matrix decomposition model are presented in this paper to detect salient objects in images.Chapter 1 briefly introduces the background of image salient object detection and the research status at home and abroad.We also expatiate the contribution and basic structure arrangement of this article.Chapter 2 summarizes relevant theoretical knowledge of image salient object detection.In addition,we introduce some common test data sets and relevant evaluation metrics of salient object detection.Chapter 3 proposes a non-convex low-rank structured matrix decomposition model for salient object detection.Since solving the nuclear norm minimization problem usually leads to a sub-optimal solution.We adopt the Bayesian view and obtain a non-convex low-rank constraint by using the maximum posterior estimation?MAP?.The new low-rank optimization model has a closed form solution.And the decomposed matrix L is located in a lower rank subspace with a cleaner background.Chapter 4 presents a novel sparse gradient based structured matrix decomposition model for salient object detection.Traditional Laplace regularization is replaced by sparse gradient regularization?i.e.,the₁l-norm of the gradient of the sparse matrix S?in the model to avoid excessive smoothness in matrix decomposition and reduce the correlation between low-rank matrix and structured sparse matrix,so as to achieve better separation effects.Chapter 5 elaborates the idea that applying tensor method to video salient object detection in future work,and preliminarily establishes the framework of video salient object detection model based on tensor decomposition.