Research On Semantic Segmentation Of Scene Image Based On Deep Learning

Image semantic segmentation technology is popular on the road of computer vision development.It has broad development prospects and application value.Image semantic segmentation started in the 1960 s,when researchers proposed a variety of image semantic segmentation algorithms.However,these traditional semantic segmentation algorithms can only segment simple images through shallow features,and the segmented images have problems of high subjectivity and low accuracy.The deep learning algorithm can predict and classify each pixel in the image,which greatly enhances the accuracy of the semantic segmentation algorithm.Based on the deep learning method,thesisr designs two kinds of crawling methods that can capture different features and integrate multiple feature information.The important work of thesis is as follows:(1)Thesis improves the problem that the neural network can not fully utilize the detailed information and affect the accuracy of the segmentation result.Firstly,according to the idea of integration,the idea of combining multi-feature fusion with neural network is proposed.Then,the existing network and related techniques of semantic segmentation are deeply analyzed,and feasible improvement points are proposed.(2)Thesis proposes a multipath residual block improved by the characteristics of the Res Net network,superimposing small-sized convolution blocks in the residual block and adding residual branches.This allows the improved residual algorithm to improve spatial detail and make better use of the details.(3)Thesis proposes parallel atrous convolution to capture multi-scale feature information.Unlike the residual algorithm,the convolution block is no longer superimposed.Parallel hole convolution has the characteristics of a large receptive field to more effectively capture more context information.Finally,the two improved methods are experimentally verified on the Cam Vid dataset and Cityscapes data,and the segmented results are compared with other networks to verify the validity of the network.