Research On Image Semantic Segmentation Based On Edge Perception

Image semantic segmentation is a basic task in computer vision,and it is widely used.Most existing methods generate dense predictions for semantic segmentation by acquiring rich context information and aggregating features within the class.However,the method based on context aggregation ignores the sparse pixels and insignificant features in the edge area,resulting in poor segmentation of the edge area.There have been some methods to solve the problem of inaccurate segmentation of edge regions by emphasizing edge information,such as using attention mechanisms,additional edge assistance,and using edges as additional categories.However,these methods usually require a large increase in computational expenses,lack of effective interaction or insufficient relevance for the assistance of semantic segmentation.This paper proposes an image semantic segmentation method based on edge perception,modeling semantic segmentation and edge detection tasks separately,using more relevant semantic edge detection tasks as auxiliary tasks,and innovatively using confidence probability maps to fuse information from the two types of tasks,To assign different fusion intensities to different areas of the image to reduce the influence of significant internal features of objects on the features of edge regions.In addition,the edge information is inferred from the semantic segmentation information,and then combined with the prediction results output by the edge detection task,and the corresponding loss is calculated for the combined edge information.In this way,on the one hand,the internal consistency constraints of the task are increased,on the other hand,To help the semantic edge detection task suppress nonedge noise and obtain higher quality auxiliary information,thereby improving the overall semantic segmentation performance.In this paper,experiments are conducted on multiple public data sets.The experimental results show that the work of this paper can improve the segmentation effect of edge regions and enhance the results of image semantic segmentation.