An Image Classification Method Based on Sequential Multi-source Domain Adaption
Lan Wu, Han Wang, Jiawei Fan, Yu Shao · 2023
There are quantity differences among multi-source domains and data distribution differences between source domain and target domain in multi-source domain adaptation, which makes it difficult to extract transferability features among all domains and makes the classification performance of the target domain poor. To address this problem, this paper proposes an image classification method based on sequential multi-source domain adaption method (SMSDA). A source domain arrangement mechanism is proposed, which takes the distribution differences and quantity differences among domains as description metrics. The target domain can perform domain adaptation with the source domain sequentially, which solves the problem that the classification performance varies greatly under different domain adaption orders of the source domain and target domain. Meanwhile, SMSDA proposes a multi domains feature local alignment rule through the introduction of fine-grained idea. It perfects the low transfer performance in the global distribution alignment. Furthermore, an adaptive adjustment strategy of sample weight is designed to improve the low classification accuracy under class-imbalance. Finally, the experimental results on three benchmark datasets show that the proposed method performs better in classification accuracy compared with the conventional methods.