Semantic Segmentation Of Urban Streetscape Images Based On PSPNet Methodology Research

In recent years,with the continuous development of artificial intelligence,machine vision,sensors and other technologies,autonomous driving and intelligent traffic supervision systems have gradually entered people’s daily lives.The correct semantic segmentation results of streetscape can help vehicles and traffic facilities to identify road,vehicles,pedestrians,traffic signs and other elements,so as to achieve effective traffic flow planning and management.However,the process of street view semantic segmentation faces many obstacles such as strong image illumination changes,complex backgrounds and serious occlusion,which also leads to the final segmentation results will be unclear edge segmentation of things,small things wrongly divided and omitted and other problems.Therefore,this thesis proposes a semantic segmentation model for urban street scenes based on improved PSPNet to improve the segmentation accuracy of street scenes based on deep learning image semantic segmentation techniques,and designs a multiscale residual network image enhancement algorithm that fuses edge features to enhance the contour features in street scenes.The main research is as follows:(1)A semantic segmentation method of urban streetscape images based on convolutional neural networks is studied.Focusing on the feature extraction principle of urban streetscape images,the training process of image semantic segmentation model based on deep learning is analyzed,and the effect of pre-processing methods on streetscape image data is compared.Finally,the PSPNet network model is applied to the semantic segmentation task of streetscape images,and its backbone feature extraction structure and enhanced feature extraction structure are analyzed.(2)The method of constructing the semantic segmentation model of streetscape is studied.The problem of imprecise edge segmentation and unclear segmentation of fine things is easily caused by classical semantic segmentation methods in the process of streetscape semantic segmentation task.In this thesis,we propose an E-PSPNet algorithm network that combines semantic and edge features.By embedding a parallel CBAM attention mechanism module in the semantic segmentation sub-network,effective features are acquired more efficiently;by adding an edge detection sub-network,visually significant boundary contours are extracted from images to obtain more accurate edge features;a feature fusion module is designed to achieve the fusion of semantic and edge features through feature stitching by Concat,thus improving the semantic segmentation accuracy.(3)The edge enhancement method of streetscape images is studied.A super-resolution reconstruction method combining edge gradient information is designed to enhance the edge feature information in streetscape images and provide high-quality data sets for model training to achieve better semantic segmentation results.Firstly,we fuse the gradient edge information of streetscape images as global features,and then densely concatenate ten multi-scale residual blocks for deep feature extraction of streetscape images,while effectively reducing the number of network parameters,and finally obtain a high quality streetscape dataset by image reconstruction module.Cityscapes city street scene data are used as the experimental dataset to investigate the effectiveness of the proposed E-PSPNet model network.The experimental results show that the pixel accuracy PA,the category average pixel accuracy MPA,and the average intersection ratio MIou evaluation indexes of the E-PSPNet model reach 97.21%,86.52%,and 78.59%,respectively,on the Cityscapes dataset,and the segmentation of the edges of things and the capture of fine things in its visualization segmentation results are significantly optimized.In summary,the E-PSPNet model proposed in this thesis has been significantly improved compared with other classical semantic segmentation models,both in terms of classification accuracy and segmentation effect,proving that the E-PSPNet model can well meet the needs of semantic segmentation of urban streetscape images.