Research On Domain Adaptation Methods For Open Class Recongnition

The purpose of Unsupervised Domain Adaptation(UDA)is to use label-rich source domain data to help learn target domain data without labels.Such methods require source domain data with correct labels.However,in reality,due to privacy or security issues,may be only the model interface after the source domain completed training is available,and only pseudo-labels obtained from the output of the source domain model can be obtained by inputting target domain data.This type of domain adaptation scenario is called black-box domain adaptation.Black-box domain adaptation usually has only one source domain,but in real tasks,there are usually multiple source domain model interfaces,and there is a problem of label shift between the source and target domains,that is,the category sets in the target domain and the source domain are not completely consistent.The shared categories are known categories,and the non-shared categories are private categories.Furthermore,when the target domain data lacks unknown samples but has descriptive information,it is hoped that unknown categories can be classified.This is called zero-shot learning(Zero-shot Learning).Therefore,this article studies domain adaptation methods for the above scenarios,mainly including the following two parts:First,a universal domain adaptation method from multiple black-box sources called Um2B(Universal Domain Adaptation from Multiple Black-Box Sources)is proposed.This method first proposes an adaptive domain attention mechanism to seek the best combination of source domain interfaces,whose performance is not worse than a single interface.Then,in order to solve crossdomain label drift,an adaptive class attention mechanism is used in adaptation to suppress the influence of private classes in the source domain.Finally,unknown classes in the target domain are identified by corrected pseudo-labels and separated from known classes.Experiments show that this method is not worse than the best single source and has better discrimination for unknown classes under the condition that the average classification accuracy of benchmark datasets does not decrease.Second,a zero-shot feature classification dual correction domain adaptation method called FCDC(Feature Classification Dual Correction)is proposed.Most existing zero-shot domain adaptation methods directly use feature extraction models trained separately on Image Net and ignore crossdataset biases between Image Net and Generalized Zero-Shot Learning(GZSL)benchmarks.This bias inevitably leads to poor visual feature quality for GZSL tasks.To this end,a simple and effective GZSL method is proposed that uses a semantic feature refinement module to generate visual maps of models uniformly and refine visual features of known and unknown class samples.In addition,most existing methods generate a large number of unknown class samples,which causes classifiers to overfit on some features.Therefore,logit adjustments are made to classifiers to improve generalization performance.