Research On Unsupervised Domain Adaptation Image Classification Method Based On Adversarial Learning

In actual image classification applications,due to the difference in weather conditions,lighting,and image acquisition equipment,there are differences in the distribution of images with the same semantic content,significantly limiting the application of image classification algorithms.However,collecting a large amount of labeled data is time-consuming and laborintensive,which promotes the research of unsupervised domain adaptation algorithms.Unsupervised domain adaptation(UDA)aims to transfer knowledge from a source domain with many labels to a target domain without labels.Current research mainly focuses on unifying the feature distributions of the two domains.However,the target domain usually has more complex background information,the global feature distributions of the source and target domains are different,and directly sharing the entire network between the source and target domains to enforce global distribution alignment will lead to performance degradation.Moreover,methods based on global feature distribution alignment often ignore the inherent discriminative information of the target data,leading to misclassification.Class alignment-based methods enhance the discriminating features of objects by generating pseudo-labels for target domain samples.However,these pseudo-labels are usually too noisy and thus inevitably affect the performance of UDA.This paper researches the above problems,and proposes two new UDA frameworks based on confrontational learning.The main research work and achievements are as follows:(1)A novel adversarial domain adaptation model is proposed based on domain-specific batch normalization.First,using the adversarial learning loss module,considering domain alignment and class alignment,a new loss function is automatically constructed from the confusion matrix obtained by adversarial learning to correct the pseudo-labels in self-training;Second,a domain-specific batch normalization module(DSBN)is introduced in the encoder architecture of a convolutional neural network(CNN),which adapts to the source and target domains separately by separating the batch normalization layers.Separate domain-specific information from domain-invariant information and learn domain-invariant feature representations to achieve better generalization performance.Finally,the accuracy of the method in this paper in the Office-Home dataset and Office-31 dataset reached 67.4% and89.4%,respectively,which verified the model’s effectiveness.(2)Propose an adversarial domain adaptation model(Swin UDA)based on feature space constraints.First,for domain alignment,the Swin Transformer is combined with adversarial adaptation to improve the robustness of the model to noisy inputs.The experimental results show that using the transformer as a feature extractor has higher transferability.Second,constraints are directly enforced in the feature space for class alignment using Orthogonal Projection Loss(OPL).Samples from the same class(whether from the source or target domain)are pulled closer,while samples from different classes are pushed away.In addition,the orthogonal projection loss is more robust to the influence of label noise interference.To preserve the discriminative information of the target domain,a mutual information maximization loss(IML)is introduced to protect the discriminating features of the target domain.The Swin UDA model proposed in this paper can simultaneously learn transferable and differentiable features.Experiments were performed on the three public datasets Office-Home,Office-31,and VisDA-2017.SwinUDA showed the best performance.