The Image Semantic Segmentation Algorithm Based On Improved Residual Network And Multiple Feature Association Module

Computer vision is indispensable in various fields such as automatic driving and intelligent security,and image segmentation is an important research direction in the field of computer vision.With the development of deep neural network technology,the image semantic segmentation task based on neural network has ushered in a new leap.Semantic segmentation of images is an important link for computers to obtain image information by recognizing image content,and provides computer-recognizable information for subsequent image processing.Its accuracy directly affects the subsequent image processing effects.Based on the framework of deep neural network,we study the task of semantic segmentation of urban street scene images.The current street scene image segmentation task has the problems of uneven training samples,inaccurate segmentation of small objects,easy segmentation errors of details such as edges,and relatively simple supervision methods.Based on the current problems in this direction,this article starts from three links.First of all,in the research of image feature extraction,we propose a feature extraction network based on an improved residual network to expand the receptive field of the network and obtain more semantic information.At the back end of the residual network,a parallel dilated convolution module is designed to obtain the image features of targets of different scales and prepare for the subsequent image feature fusion.Secondly,in the research of image feature fusion,this paper proposes a multiple feature association module.With the advantage of the attention mechanism,this part can merge the semantic information of two dimensions of different pixels and different features.Then,in the training method of the network,this paper adopts the combined training method of the main loss function and the auxiliary loss function,and introduces the Focal Loss to alleviate the problems caused by the uneven distribution of the training samples.In addition,we also design a new loss function based on the Focal Loss,and theoretically analyze its effectiveness.Finally,combined with the image semantic segmentation algorithm based on improved residual network and multi-feature correlation module proposed in this paper,experimental verification and comparative analysis are carried out,and the feasibility and effectiveness of the algorithm proposed in this paper are verified through experiments.