Research On Visual Unsupervised Domain Adaptation Methods

With the explosive growth of visual data on the Internet,the importance of computer vision-related research has become increasingly prominent.Due to the booming development of deep learning technology,computer vision methods based on convolutional neural networks and Transformers have made tremendous progress.However,deep models rely heavily on large-scale annotated datasets and suffer from poor generalization capabilities.When a deep model is deployed in an environment that is very different from the training data,the performance will drop significantly.If you train a new model for each new deployment environment,you will need to spend a lot of labor costs to label the data and hardware costs to retrain the model.Unsupervised domain adaptation solves this problem by transferring knowledge from a labeled source dataset to an unlabeled target dataset with different distribution.Traditional unsupervised domain adaptation methods face the following three challenges: firstly,the decision boundary information between classes is not considered during feature alignment,which can easily lead to negative transfer;secondly,they only consider the closed-set scenario where the source and target label sets are the same,and cannot handle the category-shift scenario where the label sets between domains are not exactly the same;thirdly,it is assumed that the model can access the source data and target data at the same time during domain adaptation,and cannot handle the source-free scenario where access to the source data is restricted.This dissertation will conduct research on visual unsupervised domain adaptation from the above three aspects.The main work and contributions of this dissertation are summarized as follows:(1)In the closed-set scenario,to address the problem that traditional adversarial learning methods ignore inter-class decision boundaries when aligning features,this dissertation proposes a closed-set unsupervised domain adaptation method based on adversarial learning and interpolation consistency,called ALIC.ALIC generates robust and informative pseudo-labels for unlabeled target samples through prediction average and label sharpening techniques,and introduces interpolation consistency into adversarial domain adaptation to optimize decision boundaries more efficiently.Since ALIC considers both feature alignment between domains and decision boundary optimization between classes,this method can learn features that are both domain invariant and target discriminative.Experiments were conducted on handwritten digit recognition and object recognition datasets to verify the effectiveness of ALIC.(2)In the category-shift scenario,in order to solve the problem of domain gaps causing interference to the separation of known classes and unknown classes,this dissertation proposes an open-set unsupervised domain adaptation method based on deconfounding domain gaps,called OSDDP.OSSDP designs a module of deconfounding domain gaps,which can reduce confusion caused by domain gaps when the model performs separation of known classes and unknown classes and domain adaptation.The mechanism of deconfounding domain gaps is to transfer the image style and context information of the target domain to the source domain,so that the model can determine whether a sample is a known class or an unknown class based on category semantics,rather than due to the confusion caused by wrong image style or contextual information.Furthermore,OSDDP designs an ensembling multiple transformations module to generate calibrated recognition scores for target samples.The experiments were conducted on two commonly used datasets and verified the effectiveness of OSDDP.Especially in identifying unknown class samples,OSDDP showed obvious advantages.(3)In the scenario where category-shift and source-free exist at the same time,to address the problem that existing white-box source-free unsupervised domain adaptation methods can only be used in a single setting,this dissertation proposes a universal white-box source-free unsupervised domain adaptation method based on style-augmented open-set consistency,called UWBSF.UWBSF uses a source white-box model to provide supervision information for the target domain,and is a general method capable of handling open set unsupervised domain adaptation and open partial set unsupervised domain adaptation.The training process of UWBSF includes two stages: source domain model generation and model adaptation.For source model generation,UWBSF trains a closed-set classifier and an open-set classifier,so that the source model has the ability to classify shared class samples and detect unknown class samples.For model adaptation,UWBSF designs an intra-domain style augmentation strategy to generate style-augmented target images,and a style-augmented open set consistency loss function to minimize the impact of target domain image style changes on model behavior.Experiments are conducted on four commonly used datasets to verify the effectiveness of UWBSF.(4)In the scenario where both category-shift and source-free exist,to solve the problem that existing black-box source-free unsupervised domain adaptation cannot handle the open-set setting,this dissertation proposes a new open-set blackbox source-free unsupervised domain adaptation setting.To address this task,this dissertation proposes an open-set black-box source-free unsupervised domain adaptation method based on knowledge distillation,called OKRA.OKRA uses an open-set knowledge distillation framework to learn the output of the black-box predictor,which can simultaneously achieve knowledge transfer of known classes and identification of unknown classes.In order to reduce the interference of noisy predictions of black-box models,OKRA employs the neighborhood similarity regularization that can utilize local structural information of the target domain.Furthermore,OKRA designs an energy-based uncertainty modeling strategy that can effectively distinguish between known and unknown target data without any threshold.Experiments were conducted on object recognition and remote sensing image cross-scene classification datasets,verifying the effectiveness of OKRA.