Research On Lightweight Semantic Segmentation Algorithm Based On Semantic Relocation

In recent years,with the rapid development of artificial intelligence technology and deep learning methods,drones and auto-driving cars emerged from scratch,and received intensive attentions from various fields.The task of semantic segmentation plays a key role in the field of autopilot.Compared with traditional semantic segmentation methods,semantic segmentation methods based on deep learning can better meet people's requirements for intelligent computing in the field of autonomous driving.The task of semantic segmentation is to perform pixel-level classification and prediction of images.The difficulty lies in the accurate prediction of the pixel-level and edge division of similar objects.To capture the global information of the segmented object,most of the existing methods adopt a network model based on the code-structure to quickly expand the receptive field of the network.However,continuous down-sampling caused irreversible loss to the spatial information of the feature map.Aiming at the semantic segmentation task in the street scene road scene,the main research work of this paper is as follows:(1)This thesis proposes SRPNet(Semantic Relocation Parallel Network for Semantic Segmentation).Specifically,we designed a high-resolution path(Global Spatial Path,GSP)to extract rich spatial information while maintaining high resolution.In another feature extraction path,we use a powerful feature extractor to expand the receptive field of the network through fast down-sampling.(2)Also,we designed a Semantic Relocation Module(SRM)based on the semantic category of objects to compensate for the lack of contextual information caused by multiple down-sampling.We use Dice Loss to alleviate the imbalance of positive and negative samples in the data to obtain better segmentation performance.(3)Finally,we tested performance of proposed method on the Cam Vid dataset and Cityscapes,two authoritative datasets.The experimental results show that SRPNet is conducive to learning the features of slender objects in the image,and can integrate high and low-dimensional features well.Compared with the existing mainstream real-time semantic segmentation methods,SRPNet has achieved very competitive results,achieving 74% m Io U results on the publicly authoritative data set Cityscapes test set.