Research On Salient Object Detection Algorithm Based On Edge Enhancement

Salient object detection refers to detecting the most eye-catching area of an image,generating a prominent image and helping people to get useful information quickly.After processed,the calculation,the storage space and the time cost of redundant data can be reduced.Salient areas and specific objects processed by the model can represent a type of scenario,enabling computers to learn complex scene understanding problems.In recent years,most saliency detection methods utilize neural network structures to extract image features through a series of convolution operations.However,a large number of convolution and pooling operations will lost the underlying features and details of the image,resulting in incomplete detection of the salient object.To this end,this paper proposes a model of underlying feature fusion and deep edge supervision.The features of the convolution operation are integrated into the deconvolution layer through the fusion structure,and each layer deconvolution output uses a joint supervision mechanism of edge and significance.Further strengthen the edges.Experiments have shown that edge detection has a certain intrinsic connection with saliency detection.In order to make the edge detection promote the improvement of saliency detection,this paper proposes a saliency detection model based on edge detection.The whole model is divided into two parts,namely the edge detection network branch and the saliency detection network branch.Each side output of the edge detection network branch is merged into the saliency detection network branch,and edge feature fusion is performed by the edge fusion module.In addition,this paper designs a special deconvolution operation module to fuse the fusion map combined with the output of the edge fusion module and the features of the convolution operation,so that the network can fully learn the characteristics of different levels in each stage.The two algorithms presented in this paper were tested on six public databases and compared with 14 advanced algorithms for current saliency detection.The experimental results show that the performance of the two algorithms is relatively advanced.