Domain adaptive retrieval (DAR) has obtained the high-precision and fast retrieval across different domains by reducing domain discrepancies. However, existing works still face two issues: (1) incorrect pseudo labels in the target domain cause error accumulation during training, and (2) most existing methods are limited to single-label works and ignore the requirement for finer-grained retrieval. To address these problems, we propose a method called Dual Semantic Guidance and Cluster Matching for Domain Adaptive Retrieval (DSG-CM) . First, samples with high confidence in the target domain are gradually selected to reduce the impact of incorrect pseudo - labels. Second, the distinctiveness of samples is enhanced by incorporating multi-label semantic information from features with the class label information of the samples. Lastly, samples are divided into micro-clusters and the inter-cluster relationships between different domains are explored to enhance the retrieval precision.This method can generate compact and efficient hash codes, leading to superior performance.Experiment results on three benchmark datasets demonstrate that our method outperforms others.

Graph-based multi-view clustering methods are widely explored. However, there are still two problems with existing methods: 1) Although some methods divide the similarity matrix into a consistency matrix and an inconsistency matrix, it is difficult to deal with the consistency information that has been misclassified in the inconsistency matrix, resulting in insufficient extraction of the valid information; and 2) Although some methods obtain a unified similarity matrix with a block diagonal structure, they do not remove redundancy information from the unified similarity matrix. To address these two issues, this paper proposes a Multi-view Consistency Learning with Block Diagonal Guidance (MCLBDG) method. First, we obtain a similarity matrix for each view via low rank representation and adaptive neighborhood. Second, we divide the similarity matrix of each view into a consistency matrix and an inconsistency matrix. The inconsistency part of different views is sieved via Hadamard product. During iterations, the misclassified consistency part can be gradually extracted from the inconsistency information. In addition, block diagonal guidance is proposed to remove the redundancy information in the unified similarity matrix as much as possible, which reduces the interference of extra-cluster samples. Finally, spectral clustering is incorporated into the model to obtain clustering results directly. Comparative experimental results on the commonly used datasets demonstrate the superiority of the method over the existing methods.

For domain shifts, domain adaptation (DA) promotes label propagation by reducing distributional differences. In recent work, DA only enhances the discriminability of linear data by linear discriminant analysis, which ignores real-world nonlinear data. Meanwhile, these methods fail to consider the negative impact of the target low-confidence samples on the training process. Therefore, confidence sample selection and specific feature enhancement for domain adaptation (CSS-SFE) is proposed. Firstly, the framework selects target high-confidence samples through the min-max principle to reduce the impact of incorrect pseudo-label during training; Secondly, the class scatter matrix and neighbor scatter matrix are balanced to enhance the features of linear and nonlinear datasets so that the discriminability of the samples is improved; Thirdly, the framework maintains source and target specific features by learning different projection matrices, which prevents the discriminability of the samples decreasing; Fourthly, the marginal and conditional distribution alignment is further applied to reduce the domain distribution discrepancy; Finally, extensive experiments on several benchmark datasets demonstrate the superiority of CSS- SFE over state-of-the-art methods.