Research On Semantics Segmentation Algorithm Of Road Scene Based On Convolution Neural Network

In recent years,in order to greatly improve people's travel efficiency and reduce the occurrence of traffic accidents,autonomous driving has gradually become a research hotspot.In the automatic driving system,the perception of the surrounding environment is the most important,and the most common method for the perception of the road scene environment is the semantic segmentation of the road scene.With the continuous development of deep learning theory,the semantic segmentation technology based on deep convolutional neural network is replacing the traditional artificial design scheme.Therefore,this article focuses on the research of the semantic segmentation algorithm based on the deep convolutional neural network road scene,through analysis Existing problems in existing algorithms,a more accurate and fast semantic segmentation algorithm is proposed.The proposed algorithms are as follows:(1)Multi-scale input feature extraction semantic segmentation network MFNet:For the fully convolutional neural network,the features extracted when acquiring the target's semantic information are single and unrepresentative,and the semantic information recovery is not detailed enough,resulting in more similarity.The problem of large target recognition error and fuzzy edge segmentation.This thesis proposes MFNet,which uses multi-scale input to input pictures of different scales into a parallel feature extraction module to enrich the features.Diversity,enhance the identification ability of the network.At the same time,in order to make the recovered semantic information more detailed,layer-by-layer deconvolution is used for upsampling.Verification on multiple datasets shows that MFNet can effectively distinguish objects with greater similarity,and the contour of segmentation is clearer,which greatly improves the accuracy of semantic segmentation.(2)Fast semantic segmentation network RINet: Aiming at the speed of semantic segmentation,on the basis of ERFNet,a fast semantic segmentation network RINetbased on deep separable convolution is proposed.The network uses a conventional residual block and an inverted residual block as the core,and the conventional residual block can make the calculation inside the network more sufficient and ensure the segmentation accuracy.The inverse residual module combined with depth separable convolution reduces the number of parameters without sacrificing segmentation accuracy.At the same time,expanded convolution is used in the inverse residual block to expand the receptive field and extract rich semantics.information.Verification on multiple data sets shows that the fast semantic segmentation algorithm RINet proposed in this thesis can improve segmentation efficiency without sacrificing segmentation accuracy,and achieves a good balance between speed and accuracy.