Research On Unsupervised Domain Adaptive Semantic Segmentation Of Remote Sensing Images Based On Dense Context

Semantic segmentation aims at pixel-level classification of images,assigning a semantic label to each pixel in an image,and it has been widely used in many fields such as self-driving,bioinformatics,and remote sensing.The quality of remote sensing images is highly susceptible to factors such as weather conditions,lighting conditions,feature patterns,feature styles and shooting angles,which also lead to a large number of remote sensing image datasets with high intra-class variability,inter-class similarity and multi-scale characteristics,and it is difficult for existing semantic segmentation methods to ensure intra-class consistency while leaving sufficient detail information.In addition,deep convolutional neural network methods usually require collecting a large amount of time-consuming and laborious labeled data for supervised learning,and the recognition results of such models are usually unsatisfactory when there is a domain gap between the training image(source domain)and the test image(target domain).The unsupervised domain adaptation method can solve the difficulty of labeling samples and enhance the generalization ability of the model by reducing the feature gap between the source and target domains.In this thesis,the unsupervised domain adaptation problem in semantic segmentation of remote sensing images is investigated,and the main work and contributions include the following aspects:(1)An unsupervised remote sensing image domain adaptation semantic segmentation network based on self-training and adversarial training is proposed.Recent research results show that the composite architecture based on adversarial training and self-training achieves leading performance in the semantic segmentation domain adaptation task,but performs poorly in remote sensing scenarios.This is limited by the intra-class variability,inter-class similarity and pseudo-label reliability of remote sensing data.Based on the above methods,this study proposes a remote sensing unsupervised domain adaptation semantic segmentation network based on self-training and adversarial training,where using inter-class separability can improve pseudo-label quality and combining dense contexts can improve the prediction accuracy of the model.Experiments on several datasets show that the proposed method can effectively improve the performance of remote sensing domain-adaptive semantic segmentation.(2)A pseudo-label generation strategy based on inter-class separability is proposed.There are a large number of objects with similar features in remote sensing images,and the confidence of model prediction for them is low,while most pseudo-label generation strategies produce pseudo-labels based on the normalized prediction confidence of the model,which will lead to the low reliability of the generated pseudo-labels,and these unreliable pseudo-labels may guide the training of the model to the wrong direction.To address this problem,this work proposes a strategy for pseudo-label generation based on inter-class separability.If the inter-class separability of a pixel reaches a desired threshold,the pixel is saved as a pseudo-label.It is experimentally demonstrated that the pseudo-labeling strategy with the introduction of interclass separability can generate more reliable pseudo-labels.(3)A dense context information module is proposed.The quality of remote sensing images is highly susceptible to factors such as weather conditions,lighting conditions,feature patterns,feature styles and shooting angles,which also lead to a large number of remote sensing image datasets with high intra-class variability,inter-class similarity and multi-scale characteristics.However,existing semantic segmentation models are difficult to retain sufficient detail information while ensuring intra-class consistency.To address this problem,this thesis proposes a dense contexts information module,which organizes each context representation sub-module of different scales in a cascading manner,and modeling the relationship between pixels by introducing multi-scale context information,which effectively improves the model’s ability to perceive the semantic environment in which multi-scale objects are located.