Research On Image Semantic Segmentation Based On Improved U-net Model

Image semantic segmentation is the most basic tasks in computer vision,and plays a key role in image understanding.It aims at the segmentation of the U-net model in the encoder-decoder framework.There is pixel information loss in the semantics segmentation.For the problems of low segmentation accuracy and unclear boundaries,a deep multi-branch residual model fusing inverse gated attention is designed.Firstly,in order to solve the problem of decreased segmentation accuracy and slow convergence speed and inaccurate classification caused by the increase in the depth of the neural network,a deep residual learning module is designed to improve the training efficiency and convergence speed of the network.Then,a deep multi-branch residual module is proposed in order to enrich the boundary information of the image,which uses convolution parallel method to extract contextual feature information at multiple scales.Finally,in order to accurately restore the up-sampling semantic features and improve the accuracy of semantic segmentation,an inverse gating-control module is designed in the first four layers of the skip connections layer,and the expression of important feature information is enhanced through counterclockwise horizontal transmission,which enriches the diversity of up-sampling information.Aiming at the problem of difficult to obtain fine features from low spatial resolution images in the model,multi-scale heterogeneous pyramid attention is used at the upper level of the encoder to capture high-level context information at different scales to improve the accuracy of segmentation.Experiments on the Cam Vid datasets show that PA of IG-DMRN model for image semantic segmentation is 91.80%,and the loss is reduced to 0.21.Compared with the original U-net model,the verification set accuracy is increased by 15.0%,and the loss is reduced by 1.22,it is better than the current image semantic segmentation method in performance.Model integrates the U structure of the original U-net model and the characteristics of high and low layer fusion,which can effectively improve the segmentation efficiency of the model,enrich the detailed features of the target boundary,realize the accurate segmentation of the target category.There are 33 figures,12 tables and 65 references in this paper.