Real-time Semantic Segmentation In Street Scene Based On Double Branch Structure

Real-time semantic segmentation is a computer vision technology widely used in fields such as autonomous driving.Achieving a balance between accuracy and speed is the difficulty of real-time semantic segmentation.This thesis focuses on real-time semantic segmentation in street scenes and the main work is as follows:(1)A Multi-Scale Large Kernel Convolution module was designed to extract features of different receptive field sizes using large kernel convolution and convolution kernels of different scales.Convolution decomposition and Depthwise convolution techniques were adopted to reduce the number of parameters for lightweight implementation.Based on this,a Multi-Scale Large Kernel Convolutional Network was proposed for real-time semantic segmentation,which can better solve the problem of feature extraction efficiency.Compared with classification networks and other backbones used for real-time semantic segmentation,this network has a larger receptive field size and aggregates features of different sizes of receptive fields at different stages.(2)A real-time semantic segmentation network with a dual-branch structure was proposed to solve the difficulties of lightweighting and feature fusion in dual-branch structures.A Strip Pooling Module was added to the detail branch to further extract spatial detail features,while a Multi-Scale Large Kernel Convolutional Network was used in the semantic branch,and channel pruning was adopted to avoid redundancy in feature extraction.In addition,a dual attention fusion layer was proposed to fully fuse the different features extracted by the detail and semantic branches.Experimental results showed that the real-time semantic segmentation network proposed in this thesis achieved a better balance between accuracy and speed.