Salient Object Detection Algorithm Based On Deep Learning

The salient object detection of the image is to solve the problem that computer vision simulates the vision of the human eye,that is,the region that the human eye notices first is detected.Salient object detection aims to highlight visually significant target areas in an image.This method is widely used in various fields such as medical,sports,and artificial intelligence.In this paper,based on a systematic study of a variety of salient target detection based on deep learning,it is found that a fully convolutional network(FCNs)has a greater advantage in conspicuous target detection tasks than a convolutional network(CNNs).However,most FCNs methods have salient object edge information loss.In order to solve this problem,the paper proposes an improved model that extracts salient object edge information features and merges multiple levels of salient object features.First,the short connection structure of the model network is removed,and the edge output images of the convolutional layers at various stages of the network are extracted.The shallow edge output has rich edge information features,while the deeper edge output loses the edge information of the salient object.But it highlights the location information features of salient objects.The image alpha matting is used to further refine the salient object map,the experimental results show that the salient target edge features of some images can be extracted,but it is less robust to images with similar foreground and background colors.For this purpose,an improved U-Net network is used to segment the shallow edge output with significant object edge refinement.The model makes full use of the object features of different sizes,and extracts the salient object edge features and the salient object spatial location features under different edge outputs.Finally,the salient edge feature and the salient object position feature are coupled to obtain the final salient object region.Experimental results show that this method has better performance than the existing methods on 6 widely used data sets without any pre-processing and post-processing.