Unsupervised Domain Adaptation For Semantic Segmentation Of Urban Scenes Based On Lightweight Networks

Semantic segmentation,one of the key technologies for scene understanding of autonomous vehicles,can ensure the driving safety of autonomous vehicles.This method is superior to non-visual scene understanding techniques in functionality and economy.The deep convolutional neural network has a good ability to understand and express the complex and changeable street environment by extracting the deep features of the image.However,it is limited by the low operation speed of traditional deep neural networks and the need for a large amount of training data with accurate labels.Considering the above issues,this thesis has done the following:(1)A lightweight semantic segmentation method integrating multi-level spatial features is proposed.Aiming at the problem of image information loss caused by downsampling in the network operation process,based on the improved ENet,a reverse residual structure is used in the downsampling layer to expand image information and reduce the loss of spatial feature information in the image.Use spatial attention to process spatial feature information,enhance relevant features and weaken irrelevant features.Through the fusion of multi-level shallow space and abstract deep semantic features,the network’s ability to process image detail features is improved.Experiments show that the proposed method has high performance on mobile terminals such as NVIDIA Jetson TX2,NVIDIA Jetson Xavier NX and NVIDIA Jetson Xavier AGX.(2)An unsupervised domain adaptation method based on Euclidean norm and an imagelevel fine-grained class balancing method are proposed.By analyzing the relationship between the distance between classes and the uncertainty of prediction results,it is proposed to use Euclidean norm to represent the uncertainty value of prediction results.Aiming at the problem of excessively large gradients of high-confidence categories in the minimum information entropy method,a maximum Euclidean norm loss is proposed,which effectively increases the network’s processing capacity for low-probability categories.In addition,an image-level finegrained class balancing method is proposed that can effectively alleviate the class imbalance problem in the target domain.The performance of the proposed method is verified on two synthetic-to-real domain adaptation tasks,respectively,and it is improved by 5.7% and 2.9%compared with other SOTA methods under two backbone networks of VGG-16 and Res Net-101.Experiments show that the proposed method outperforms other SOTA methods in performance.(3)A lightweight unsupervised domain-adaptive semantic segmentation method is proposed.The lightweight network is improved into a dual-branch network structure,in which the semantic information branch is composed of the encoder of the original lightweight network,and a spatial detail branch with shallow structure and wide channels is designed.Finally,the two branches are combined using a bidirectional guided aggregation method.feature map fusion.The above methods are tested,and the experiments show that the dual-branch lightweight semantic segmentation network has better performance,and the proposed lightweight unsupervised domain-adaptive semantic segmentation method has certain advantages compared with other methods in this thesis.