Unsupervised domain adaptive method for beneficial feature alignment under class condition

Abstract

The invention discloses an unsupervised domain adaptive method for beneficial feature alignment under class conditions, and the method comprises the steps of calculating all source domain images and target domain images, and obtaining a pseudo tag of the target domain image; decoupling the pseudo labels of the source domain image and the target domain image through variational information bottlenecks, filtering features irrelevant to tasks out, and obtaining beneficial and migratable features; estimating intra-class difference and inter-class spacing by using a conditional slice Wharisstein distance, minimizing the intra-class difference and maximizing the inter-class spacing in a cross-domain manner, reducing class-level distribution difference between a source domain and a target domain, and obtaining domain-invariant discriminant features. According to the method, decoupling of a source domain and a target domain can be achieved, class-level information is embedded into the slice Weiisstein distance, beneficial feature alignment is achieved, and meanwhile beneficial feature decoupling and class condition feature alignment are achieved so as to promote better domain adaptation.

Description

technical field [0001] The invention belongs to the field of unsupervised domain self-adaptation, in particular to an unsupervised domain self-adaptive method for beneficial feature alignment under class conditions. Background technique [0002] Deep Neural Networks (DNNs) have achieved significant progress in various tasks, such as image classification, object detection, image segmentation, face recognition, etc. However, these impressive advances depend on the strict assumption that large amounts of well-labeled data are available for model learning in domains of interest. Manual labeling is often costly and labor-intensive; especially for data-sensitive domains such as medical images and industrial inspections, labeled samples are not even available. [0003] A general strategy (such as transfer learning) operates by reusing knowledge / models learned from available related domains (called source domains) into domains of interest (called target domains). Unfortunately, th...

AI Technical Summary

Problems solved by technology

Source and target domains basically overlap in task-related information, while redundant information from task-independent factors (such as background, color, and context) may be different in nature, and forcibly aligning these useless features may affect adaptation performance

In addition, in some existing UDA methods, there is another bottleneck: that is, the class-level distribution differences are not fully exploited during the adaptation process, and the distribution differences are only adapted at the domain level, while Differences that do not encode class-level information will make the learned features domain-invariant but class-indistinguishable, which can lead to misclassification, yielding suboptimal performance

Images