Research And Improvement Of Saliency Detection

Saliency detection is part of the image processing work and is the pre-processing step for most of its tasks.The purpose is to detect and segment the most noticeable part of the entire image.Generally speaking,this part of the image contains the largest amount of information and the most important.Today,saliency detection has been widely used in image retrieval,image classification,target detection,image segmentation and other fields.With the advancement of technology,the traditional saliency detection method based on manual feature extraction has become more mature;and the introduction of deep learning methods has also made the saliency detection efficiency further.However,the two methods have shortcomings,such as the traditional methods that rely heavily on boundaries and are incapable of dealing with complex background situations.The deep learning methods cannot fully utilize image features and the target boundaries are not obvious.Aiming at the problem that traditional methods rely too much on borders and have poor detection effect on images with small background differences,this thesis proposes a saliency detection method combining background prior and foreground prior.Unlike other methods that combine prior background priors,this thesis takes into account target features and background features,and uses a Markov absorption chain-based saliency detection method to extract background prior saliency maps.Further,in order to get rid of relying on the boundary to locate the target,a color-enhanced harris corner detection method is introduced to determine the position of the convex hull to locate the salient target.Aiming at the problems of rough image fusion and target edge blurring in deep learning methods,this thesis proposes a new feature fusion network to make full use of high-level and low-level image features and use a mixed loss function to improve edge clarity.In this thesis,the features extracted from each network layer are refined by cross-layer aggregation and the residual refinement module.The significant prediction map after refinement is continuously refined by the loop module,so that the lowlevel detail information and high-level semantic information are reasonable Use and fuse the saliency prediction map of each layer to get the final saliency map.Binary cross-entropy loss function,structural similarity loss function,IOU loss function are fused into a mixed loss function,which is used to replace the commonly used single-binary cross-entropy loss function,multi-directional training significantly predicts the error between the prediction map and the manual annotation map,thus Improve target edge clarity.Based on multiple international public data sets,this article tests the above two methods,and compares them with the mainstream saliency detection algorithms in the field to conduct experiments and analysis to verify the advancement of the two methods.Experimental results show that both methods have achieved excellent results and performance has been improved compared to other mainstream algorithms.