| With the technological iteration and update of science and the rise of the concept of smart city,street scene understanding is the premise of smart city applications such as autonomous driving.In the urban street scene image,the semantic segmentation technology with both precision and real-time is the guarantee of assisting vehicles to carry people safely,so the image processing of urban street scene is an significant research field.Based on this background,this paper selects the classic DeepLabV3+ model in the field of image segmentation to carry out research on street scene images.Based on the shortcomings of the model itself and the difficulties in urban street scene images,this paper optimised and improved the network,and achieved effective segmentation results.The main research work of this paper is as follows:1.Aiming at the problem that DeepLabV3+ model only from the spatial perspective extracts features and fails to make full use of the target category information,Deeplabv3+network model is optimized for semantic segmentation.Integrating category enhanced attention can enhance the advantages of pixel category correlation and the embedded channel attention can adaptively recalibrate class level information.A semantic Street View segmentation network integrating class attention mechanism is proposed,which improves the overall segmentation accuracy of the model and corrects the phenomenon of class confusion or class error segmentation.2.In view of the difficulty of target recognition caused by the phenomena of complex types,large size changes,and far and far location of street view images,In order to capture multi-scale features and reuse key features,this paper designs a multi-scale adaptive fusion Street Scene semantic segmentation algorithm.which used spatial attention module to extract shallow spatial information to fine pixel positioning and two-dimensional discrete cosine transform to obtain more information.Reasonable use of spatial attention mechanism and twodimensional discrete cosine transform to generate spatial and frequency correlation weights for each feature map.,which enhanced the adaptive fusion of multi-scale features and improved the segmentation accuracy of the network.3.Aiming at the problem that most image semantic segmentation networks cannot combine real-time performance with high precision,DeepLabV3+ model trunk is improved to reduce the number of parameters.Meanwhile,densely connected ASPP module is designed to increase the sensitivity field and improve the precision without increasing the number of parameters.Compared with the original network,the segmentation speed and accuracy are improved. |
