Research On Unsupervised Domain Adaptation In Image Recognition

Deep Neural Networks(DNNs)demonstrate significant success across a range of machine learning tasks.However,its effectiveness is highly dependent on a large amount of well-labeled training data,which is typically time-consuming and expensive to label manually.Meanwhile,another drawback of traditional deep learning is its inability to generalize to new datasets due to domain shift issues.Domain Adaptation(DA)addresses this issue by leveraging knowledge from label-rich source domains to assist relevant but label-scarce target domains.For instance,for tasks such as semantic segmentation that require fine-grained labeling,game engines can be used to rapidly generate a large number of annotated images to aid in real-world semantic segmentation.This dissertation primarily addresses the problem of single-source unsupervised domain adaptation,in which there is a source domain and a target domain,with source domain being fully labeled and the target domain being unlabeled,hoping to get a well performed model in the target domain。We combines three fundamental image recognition tasks:image classification,image semantic segmentation,and image object detection.The shortcomings of existing algorithms are analyzed in light of various practical application scenarios and tasks,and different methods are proposed.The following summarizes the primary research and contributions:·For domain adaptation in semantic segmentation,this dissertation proposes a joint adversarial learning method that fuses low-level and high-level domain discriminators.Typically,adversarial learning of the output probability space is used for tasks with structured outputs,such as semantic segmentation.This dissertation begins by proposing a joint adversarial learning method for boosting the domain discriminator in the output space by introducing domain discriminator information via low-level features.As a result,training of high-level decoders will be facilitated.Then,this dissertation proposes a weight transfer module for mitigating the trained decoder’s inherent bias toward the source domain.Specifically,the weight transfer module replaces the original decoder with a new one that is learned only in the presence of an adversarial loss and thus focuses primarily on domain divergence reduction.Extensive experiments on two widely used benchmarks demonstrate that the proposed method can achieve significant performance gains over various baseline methods,demonstrating the effectiveness in output space adaptation.·For the domain adaptation in object detection,this dissertation proposes a class prototype-based cross-domain alignment method for RPN network.Currently,the majority of object detection domain adaptation methods rely heavily on feature alignment on backbone networks or instance classifiers to improve detection model transferability.In contrast to this,this dissertation begins by highlighting the issue of domain differences in RPN.This dissertation proposes to perform feature alignment during the RPN stage in order to effectively distinguish foreground and background candidate boxes in the target domain.To be more precise,a set of learnable RPN class prototypes is constructed first,and then the RPN features are forced to be consistent with the source and target domain prototypes.Second,this dissertation employs Grad CAM to identify discriminative regions in foreground proposals and then weights them spatially to improve the discriminability of RPN features when they are aligned with prototypes.This dissertation conduct extensive experiments on a variety of cross-domain detection scenarios and find that the proposed method outperforms previous state-of-the-art methods.·For domain adaptation in image classification,this dissertation proposes an instance discrimination based contrastive learning method for low-confidence samples in the target domain.To construct reliable pseudo-labels,class prototypes,or cluster centers,the majority of current domain-adaptive learning methods rely on high-confidence samples.These methods omit a large number of low-confidence samples.This results in sub-optimal transferability,as samples with high confidence are generally more biased toward the source domain.To address this issue,this dissertation proposes a contrastive learning method for low-confidence samples that takes advantage of the target data’s full structure.To begin,this dissertation proposes that using low-confidence samples to construct positive and negative sample pairs can help avoid semantic conflict.It then proposes that re-representing the original features with classifier weights can help alleviate the inconsistency of the learned feature distribution and task discrimination.Second,this dissertation employs cross-domain Mixup in conjunction with the proposed contrastive loss in order to further reduce the cross-domain gap.Finally,the proposed method is effective and achieves state-of-the-art performance in standard unsupervised and semi-supervised domain adaptation task settings.To sum up,this dissertation conducts exploration and research on the domain adaptation problem in different vision tasks,combined with the specific characteristics of the problem.From the perspective of domain adversarial learning in the output space,the cross-domain alignment of the class features of the RPN network,and the target domain feature distribution learning,this dissertation proposes novel solutions.This dissertation involves a variety of transfer learning scenarios from virtual to real,different weather,and different urban scenes.The experimental results show that the method in this dissertation has achieved good results in improving the domain adaptability of the CNN model for different visual tasks.Significant progress has been made compared with existing methods,further reducing the dependence of the CNN model on labeled data,improving the generalization ability of the model in real scenarios,and demonstrating the practical application of this method in autonomous driving,industry 4.0,and smart cities.