Research On Domain Adaptation Methods Based On Image Classification

Recently,artificial intelligence(AI)technology,represented by deep learning,has de-veloped rapidly.However,in practice,the performance of artificial intelligence still de-pends on a large number of labeled data,and the generalization capability of AI models is not good enough.Therefore,how to reduce the dependence on labeled data and improve the generalization capability of AI model is an urgent problem to be solved in the field of artificial intelligence.Domain Adaptation(DA)is a good solution to this problem,aim-ing to exploit the knowledge learned from the source domain to boost the learning of the target domain.Starting from the loss function,learning framework and network architec-ture,this paper proposes two DA algorithms for both semi-supervised and unsupervised DA tasks:the DA algorithm based on efficient label propagation and the DA algorithm based on reciprocity normalization.Starting from the loss function and the learning framework,semi-supervised domain adaptation(SSDA)methods have demonstrated great potential in large-scale image classi-fication tasks when massive labeled data are available in the source domain but very few labeled samples are provided in the target domain.Existing solutions usually focus on fea-ture alignment between the two domains while paying little attention to the discrimination capability of learned representations in the target domain.Thus,this paper presents a novel and effective method,namely Effective Label Propagation(ELP),to tackle this problem by using effective inter-domain and intra-domain semantic information propagation.inter-domain propagation,we propose a new cycle discrepancy loss to encourage consistency of semantic information between the two domains.For intra-domain propagation,we propose an effective self-training strategy to mitigate the noises in pseudo-labeled target domain data and improve the feature discriminability in the target domain.As a general method.our ELP can be easily applied to various domain adaptation approaches and can facilitate their feature discrimination in the target domain.Experiments on the public benchmark test datasets show that ELP can not only improve the performance of mainstream semi-supervised DA methods,but also improve the performance of unsupervised DA methods.Starting from the design of network architecture,batch normalization(BN)is widely used in modern deep neural networks,which has been shown to represent the domain-related knowledge.and thus is ineffective for cross-domain tasks like unsupervised do-main adaptation(UDA).Existing BN variant methods aggregate source and target domain knowledge in the same channel in normalization module.However,the misalignment be-tween the features of corresponding channels across domains often leads to a sub-optimal transferability.In this paper,we exploit the cross-domain relation and propose a novel nor-malization method,Reciprocal Normalization(RN).Specifically,RN first presents a Re-ciprocal Compensation(RC)module to acquire the compensatory for each channel in both domains based on the cross-domain channel-wise correlation.Then RN develops a Recip-rocal Aggregation(RA)module to adaptively aggregate the feature with its cross-domain compensatory components.As an alternative to BN,RN is more suitable for UDA prob-lems and can be easily integrated into popular domain adaptation methods.Experiments on the public benchmark test datasets show that the proposed RN outperforms existing nor-malization counterparts by a large margin and helps the mainstream adaptation approaches achieve better results.