The Research And Application Of Saliency Detection Based On Sparse And Low-rank Matrix Recovery Model

With the rapid development of information technology and mobile Internet,the image and video data grow exponentially.Saliency detection is a kind of technology aims to detect the saliency objects in images by simulating human visual attention mechanisms.It has been widely used in many fields such as visual tracking,object discovery,human-computer interaction,etc.,and has been attracted widespread attention in recent years.The sparse and low rank matrix recovery model plays an important role in saliency detection researsh,but when the saliency object is located at the edge of the image or with the cluttered secens,the detection accuracy will be greatly dropped.This thesis mainly studies the saliency detection method based on the low rank matrix recovery model,and proposes an improved sparse and low rank matrix based saliency detection method withon background template extraction and structural constraint.Firstly,aim at the problem of the imprecise background template extraction method in the existing sparse and low rank matrix recovery model,a new background template extraction method based on semantic information is proposed.Above all,by calculating the connectivity and contrast between the image super-pixel blocks,the super-pixel block with the foreground object in the original background area is removed,and the SegNet semantic segmentation is used to obtain the semantic information of the image,the super pixel block belonging to the background template in the original foreground region is added to the background template.The experimental results show that the accuracy of our method is improved by extracting high-level features from the image.Secondly,aim at the problem of saliency value consistency between adjacent super-pixel blocks in the existing saliency detection model based on sparse and low rank matrix recovery is not considered,a structural constraint method between adjacent super-pixel blocks is proposed.By calculating the influence factor matrix and the confidence matrix between the super pixel blocks,a sparse and low rank matrix recovery model based on structural constraints is constructed,so that the super pixel blocks belonging to the same region have consistent saliency values in the obtained salient map,combined the extracted background template of the fusion semantic information improves the existing saliency detection model base on sparse and low rank matrix recovery,and improves the effect of saliency detection.The experimental results show that the proposed method makes the obtained saliency map smoother,has a consistently saliency object,and improves the performance of the three saliency detection indicators,which fully proves the effectiveness of the algorithm.Finally,a stylized watermarking system based on above saliency object detection is design and implemented.The result show that the system can increase the difficulty of automatic watermarking and help to prevent automatic watermark removal.