Research On Self-training Based Unsupervised Domain Adaptation Semantic Segmentation Algorithm

Semantic segmentation,as a core task in computer vision research,plays a very important role in fields such as autonomous driving and medical image diagnosis.However,the existence of domain shifts hinders the transferability of semantic segmentation capability in different scenarios.Unsupervised domain adaptation(UDA)based on self-training is an important method to alleviate the impact of domain shifts on model training and improve the generalization ability of cross-domain segmentation models.However,self-training often faces the problem of scarce data for difficult categories,and UDA models may have difficulty producing good segmentation results for difficult categories.In addition,there may be incorrect data in pseudo labels,and the problem of low utilization of target domain data also hinders the optimization of cross-domain segmentation models.Moreover,because the supervision of self-training and consistency learning is pixel-wise,UDA models cannot optimize the relationship between pixels from a global perspective.The strong dependence between pixels and contextual environment in the prediction process further weakens the generalization effect of UDA models.To address the above issues,this thesis mainly has the following contributions:First,this thesis proposes a domain adaptation method based on difficult category-aware data augmentation,which enriches the proportion of difficult category data in the target domain pseudo labels by cross-image data augmentation.Additionally,this method implements consistency constraints based on Mean-Teacher,providing other forms of supervision information for ignored regions to improve the robustness of cross-domain segmentation models.Second,this thesis proposes a contrastive learning framework suitable for UDA cross-domain segmentation,which improves the distribution structure of pixel features in the source domain feature space by "pulling closer to similar instances and pushing away from dissimilar instances",the feature hyperplane used for classification is maximally shared by the two data domains.Meanwhile,this method conducts contrastive learning in the shared region of the target domain,decoupling the dependence between pixel features and image context and providing more robust feature representations.Experiments on the benchmark tasks for UDA semantic segmentation demonstrate the effectiveness of the proposed methods in terms of experimental indicators and visualization results.