Research On Saliency Target Detection Model Based On Matrix Decomposition

Salient object detection is one of the most import branches in computer vision,which aims at extracting the most attractive regions from an image.It has been used as a preprocessing in image processing and has made the necessary preparations for the following set of visual image processing.As far,a great progress has been achieved in this area while it still exists some challenge problems.It is needed to design more effective algorithms to improve the performance of salient object detection.In this paper,we focused on studying image salient object detection via low-rank matrix recovery theory.The main contents of this paper are:At first chapter,we introduce the basic theories of the image salient object detection and briefly describe the research status of saliency detection.At second chapter,we give some theories and optimal algorithms,which the paper depend on.In addition,we also introduce some common datasets and evaluation metrics for salient object detection.At third chapter,based low-rank matrix recovery theory,we propose a non-convex matrix decomposition model for salient object detection.In this model,for the issue that background is not enough suppressed in saliency map,a non-convex Spnorm has been introduced to constraint on the background matrix,which makes the features of background in a lower rank space,so that the background of the detected saliency map is cleaner;also,for the foreground,considering the inconsistent saliency values of pixels within a same salient object,instead of ? · ?1norm,a group sparsity norm has been introduced to explore the relationship among pixels,which makes the pixels belonging to a same salient object share a same salient value,being consistently highlighted.At fourth chapter,following the third chapter,we proposed a diversity induced multiview matrix decomposition model for salient object detection.In this model,we add a Laplacian regularization term for each low-level feature to guarantee the structure of image features and to make the object smoothing,which makes the salient objects separated from the background more perfect.Additionally,we also add a diversity regularization term into this model,which explores the potential information among different features as far as possible and obtain more completely salient objects.The performance of the proposed model has been advanced,especially for images complex scene.