Manifold embedded joint geometrical and statistical alignment for visual domain adaptation
作者：
Highlights：
• A novel domain adaptation approach called MEJGSA is proposed.
• MEJGSA reduces the risk of degenerate feature transformation.
• A robust objective function is considered to minimize the divergence between the two domains.
• More accurate initial pseudo labels are generated.
• Thorough experiments like ablation study and feature visualization underline the validity of MEJGSA.
摘要
•A novel domain adaptation approach called MEJGSA is proposed.•MEJGSA reduces the risk of degenerate feature transformation.•A robust objective function is considered to minimize the divergence between the two domains.•More accurate initial pseudo labels are generated.•Thorough experiments like ablation study and feature visualization underline the validity of MEJGSA.
论文关键词：Manifold learning,Domain adaptation,Unsupervised discriminant analysis,Classification,Transfer learning
论文评审过程：Received 11 February 2022, Revised 31 August 2022, Accepted 11 September 2022, Available online 22 September 2022, Version of Record 13 October 2022.
论文官网地址：https://doi.org/10.1016/j.knosys.2022.109886
