| With the development of society and the increasing demand for cars,autonomous driving has become a research hot spot.With the research of deep learning,semantic segmentation,one of the fundamental tasks of computer vision,has developed rapidly and has been applied to remote sensing satellite image segmentation,medical image segmentation,farmland segmentation and automatic driving.This thesis studies semantic segmentation algorithm based on road scene datasets for automatic driving.There are some difficulties in semantic segmentation of road scene images:1.Automatic driving has high requirements on the processing speed of road images,and the existing segmentation methods are difficult to trade off the speed and accuracy.2.Semantic segmentation is a pixel-level task,which has high requirements on extracted features.Only using deep features cannot meet the requirements.3,Semantic segmentation datasets annotation requires a lot of work,so how to achieve better results with limited datasets is one of the research directions.In view of these difficulties,this thesis puts forward three models:(1)To make a trade off between speed and accuracy,Factorized and Regular Network(FRNet)is proposed.FRNet uses asymmetric encoding-decoding structure to improve speed while improving accuracy.Encoder fully extracts features and concise decoder structure recovers features,ultimately achieving pixel to pixel segmentation.The core module of the network is Factorized and Regular block(FR),which adopts three branches including residual branch,asymmetric factorization convolution branch and regular convolution branch.These three branches play their respective roles to make the network meet the requirements of semantic segmentation for real-time and accuracy of automatic driving.(2)In order to improve the accuracy of pixel-level semantic segmentation task by utilizing the features of different stages,based on FRNet mentioned above,two branches are added to construct a new network:Multi-Stage Feature Extraction Fusion Network(MFNet).The encoder of FRNet extracting deep abstract information is regarded as coarse-grained branch.Attention branch is introduced to assign weight to image pixels,so that the key information gets attention.Fine-grained branch is introduced in shallow network to retain shallow details.Finally,after the multi-stage feature fusion of the three branches(coarse-grained,attention and fine-grained branches)by the feature fusion module,upsampling generates segmentation images.The experiments proved that accuracy is improved.(3)In order to achieve higher accuracy on limited datasets,Semantic Segmentation via Improved Generative Adversarial Network(SSIGAN)is proposed.In the game between the discriminator and generator,the discriminator network constantly improves its discriminator ability,forcing the generator network to generate more realistic segmentation images.In the improved generative adversarial network,some of regular convolution in generator network is replaced by asymmetric factorization convolution to compensate the loss of information in the discriminator network.The original images and the images to be judged are connected as the input of the discriminator network to provide a benchmark to improve the accuracy of judgment.Experimental results show that improved generative adversarial network achieves better results than the single generator network and achieves higher accuracy on limited datasets.In this thesis,three convolutional neural networks for semantic segmentation of road scene datasets are proposed,which solve some difficulties in semantic segmentation in road scene.They can be used for scene perception and provide semantic information for automatic driving. |
