基于迁移学习和增强低秩表示的跨域图像识别方法

Transfer Learning with Enhanced Low-Rank Representation for Cross-Domain Image Recognition

摘要: 本文提出了一种基于迁移学习和增强低秩表示的跨域图像识别方法(TLELRR)，旨在解决跨域图像识别任务中因源域与目标域之间存在显著的特征分布差异及标签结构不一致所导致的性能退化问题，进而有效提升跨域图像识别的准确率与泛化能力。我们在经典低秩表示框架的基础上引入了稀疏正则化和图正则化，以增强模型的鲁棒性与判别性。其中，稀疏正则化被施加于重构系数矩阵，利用稀疏表示捕捉局部流形结构，从而促使目标域样本可通过少量源域样本进行有效重构；图正则化则基于样本间相似性构建图结构，并借助图拉普拉斯算子将局部几何信息引入低秩建模过程，从而提升模型对跨域差异的适应能力。与一系列经典的非深度迁移学习方法相比，本文提出的TLELRR框架在多个基准任务上展现出更优的性能。

Abstract: This paper proposes a cross-domain image recognition method based on transfer learning and enhanced low-rank representation (TLELRR), aimed at addressing performance degradation caused by significant feature distribution discrepancies and inconsistent label structures between the source and target domains. We augment the classic low-rank representation framework with sparse regularization and graph regularization to boost the model’s robustness and discriminative power. Specifically, sparse regularization is imposed on the reconstruction coefficient matrix, encouraging each target domain sample to be effectively represented using only a few source domain samples, which helps capture the local manifold structure and improves robustness. Meanwhile, graph regularization constructs a similarity graph based on sample relationships and incorporates the graph Laplacian to embed local geometric information into the low-rank modeling process, thereby enhancing the model’s adaptability to domain shifts. Compared with a series of classical non-deep transfer learning methods, the TLELRR framework proposed in this paper shows better performance on multiple benchmark tasks.