Domain Adaptive Image Classification Algorithm Based On Inter-class Discriminant Information And Domain Invariant Feature

In recent years,the rapid development of deep learning has led to significant breakthroughs in the integration of deep learning models including computer vision,acoustic processing,natural language processing and other fields.These successes come from the widespread use of large-scale labeled data and underpins the generalization of model.However,in practical applications,large-scale data is scarce due to the scarcity of data and the high cost of labeling.This has become a bottleneck problem for the development of deep learning.To address this issue,researchers have begun to focus on domain adaptation learning.Through domain adaptation learning,the model can transfer the learned semantic knowledge from the labeled source domain dataset to the unlabeled target domain dataset.Therefore,the model can achieve generalization on the target domain and obtain better cross-domain classification performance.This paper takes domain adaptive image classification as the carrier and explores domain adaptive methods from the perspectives of inter-class discrimination information and domaininvariant features.The specific results of the work are as follows:(1)In order to address the problems of insufficient domain generalization ability and high model complexity in existing closed set domain adaptation methods,this paper proposes a Hyper-Spherical Mapping of Features and Category Confusion Optimization Network(HCN)for closed set domain adaptation.Firstly,HCN introduces a feature hypersphere mapping,which converts the feature space from Euclidean space to hypersphere space.By normalizing feature vectors and classification weights,HCN reduces the influence of domain shift on decision boundaries.In addition,to further achieve cross-domain distribution alignment,HCN proposes a class confusion loss.Specifically,the discriminative information is used to optimize the intra-class compactness and inter-class discriminability on the target domain,reducing the prediction confusion.Moreover,an entropy loss has been introduced to boost the performance on target domain.It leverages the principles of entropy minimization and maximization to improve the confidence and ensure the independence and fairness of each prediction.Furthermore,the effectiveness of HCN is verified through theoretical analysis and experiments on multiple domain adaptation benchmark datasets.(2)In addition to closed-set domain adaptation,partial domain adaptation(PDA)is also explored in this paper.PDA is a more practical and challenging domain adaptation scenario,where the label space of the target domain is a subset of the source domain.This paper proposes a Confidence based class weight and embedding discrepancy constraint network(CEN)for partial domain adaptation.Firstly,a temperature coefficient is introduced in the prediction of the classifier.Lower temperature will make the prediction distribution over classes become sharper and pay less attention to outlier classes during training.In addition,existing methods measure the transferability of samples through the class weight.However,it suffers from the lack of robustness of the class weight.Therefore,CEN proposes a confidence based class weight.Considering the confidence and discriminative information of the target domain prediction,the class weight can reveal the probability that the category belongs to outlier classes or shared classes.To this end,the model can adaptively adjust the contribution of outlier and shared classes during training and reduce negative transfer while boost the positive knowledge transfer across domains.Moreover,CEN introduces an entropy minimization loss as additional supervision to encourage confident predictions on the target domain.To further promote the alignment of feature distribution,CEN propose an adaptive feature distribution alignment loss by considering the relationship between feature norm and the domain gap.Specifically,the expected feature norms of the source and target domain will be restricted to a dynamic threshold therefore the collisions and confusions in the feature space will be reduced.This paper evaluates the performance of CEN on multiple domain adaptation datasets and conducts sufficient ablation experiments,demonstrating that CEN achieves state-of-the-art performance in partial domain adaptation image classification tasks.The main contributions of this paper are as follows:(1)For the adaptive task in a closed-set domain,a feature mapping method is proposed to reduce the impact of domain shift on the decision boundary.By using class confusion loss and information entropy loss,the optimization target domain’s intra-class compactness and interclass discriminability are improved,while the prediction’s robustness and fairness are increased.(2)For the adaptive task in a partial-set domain,a confidence-based category weight is proposed to measure the transferability of source domain categories.Additionally,a feature alignment loss is introduced to promote the alignment of feature distributions across domains.