Research On Semantic Segmentation Of Road Scene Images Based On Deep Learning

With the development of technology and the improvement of life,cars have become a necessity for people's lives.Unsafe driving behavior may bring economic and safety problems.People hope to reduce the occurrence of dangerous situations through automatic driving technology,so the automated driving system becomes current research hotspots.In the automated driving system,how to analyze the current road conditions is a research focus.At present,the image acquisition device such as the camera is mature and can effec-tively obtain information about the current environment.With computer vision algorithms,such as semantic segmentation,the current road environment can be analyzed to help the automated driving system make judgments.This thesis mainly studies the semantic seg-mentation of road scene image based on deep learning method,and proposes three models:(1)A feature fusion semantic segmentation model based on multi-branch architec-ture is proposed.The model contains three branches,which are semantic branch,atten-tion branch and spatial information branch.Each branch learns different features and the learned features are fused by the feature fusion module to obtain better segmentation re-sults.The validity of the multi-branch architecture is verified by experiments.In addition,the auxiliary loss is used in the network to supervise the learning process,which further im-proves the segmentation quality.The model was tested on multiple datasets and compared with other methods to demonstrate the effectiveness of the proposed model.(2)A fast semantic segmentation model based on encoder-decoder architecture is pro-posed.Based on the residual bottleneck block,combined with the dilated convolution and others,new blocks are proposed,which effectively expands the receptive field of the net-work and improves the segmentation quality.Besides,an asymmetric encoder-decoder architecture is employed in the proposed model,which reduce the number of parameters and increase segmentation speed.The experimental results show that the proposed fast semantic segmentation model can inference in real time while ensuring accuracy.(3)A semantic segmentation model based on the generative adversarial network is proposed.Using the generative adversarial network to train the semantic segmentation model,a semantic segmentation network is first proposed,which is used as a generator in the generative adversarial network to generate the segmentation result.A weighted loss is proposed to ensure that the output of the generator is the segmentation map of the input.The proposed discriminative network is a classification network,which used to discrim-inate whether the input is ground truth.The performance of the segmentation network is improved by adversarial training.Experiments demonstrate that using the generative adversarial network can effectively help the segmentation network to get a good segmen-tation result.Besides,we convert the discriminative network into a fully convolutional network to analyze the details of the discriminative network.We observe that the discrim-inative network can indicate each region segmentation quality in the image,which further demonstrates that the generative adversarial network can improve the quality of segmen-tation.