Institute of Computing Technology, Chinese Academy IR
| PIC: Domain generalization by path information constraint | |
| Zhu, Jilong1,2; Zhuo, Junbao3; Wang, Shuhui1 | |
| 2025-12-01 | |
| 发表期刊 | PATTERN RECOGNITION
![]() |
| ISSN | 0031-3203 |
| 卷号 | 168页码:12 |
| 摘要 | Despite the widespread use of deep neural networks in various tasks, their weak generalization ability to out-of-distribution data has always been one of the challenges hindering their practical applications. Domain Generalization (DG) aims to address this issue by learning domain-robust features across source domains, enabling models to generalize well to unseen target domains. Although DG has received extensive attention, existing solutions often overlook the relationships between different source domains and fail to make good utilization of the information contained in the source domain data to enhance the model's generalization ability. In this paper, we propose the Path Information Constraint (PIC) method, which enhances training diversity by constructing distinct learning paths within source domains and leverages constraints between paths to capture consistent and generalizable features. Experimental results show that our method achieves competitive results on multiple challenging benchmarks. PIC outperforms the best non-ensembled method by an average of 0.7% across four benchmarks, especially a 3.3% improvement on the OfficeHome benchmark. Furthermore, ensembling PIC by averaging multiple independently trained models leads to further improvements. The code is available at: https://github.com/JilongZhu/path_information_constraint. |
| 关键词 | Domain generalization Transfer learning Path information constraint Adversarial manifold learning |
| DOI | 10.1016/j.patcog.2025.111769 |
| 收录类别 | SCI |
| 语种 | 英语 |
| 资助项目 | National Key R&D Program of China[2023YFC2508704] ; National Natural Science Foundation of China[62022083] ; National Natural Science Foundation of China[62236008] ; National Natural Science Foundation of China[62402033] |
| WOS研究方向 | Computer Science ; Engineering |
| WOS类目 | Computer Science, Artificial Intelligence ; Engineering, Electrical & Electronic |
| WOS记录号 | WOS:001501491400001 |
| 出版者 | ELSEVIER SCI LTD |
| 引用统计 | |
| 文献类型 | 期刊论文 |
| 条目标识符 | http://119.78.100.204/handle/2XEOYT63/42318 |
| 专题 | 中国科学院计算技术研究所期刊论文_英文 |
| 通讯作者 | Zhuo, Junbao |
| 作者单位 | 1.Chinese Acad Sci, Inst Comput Tech, Key Lab Intell Info Proc, Beijing, Peoples R China 2.Univ Chinese Acad Sci, Beijing, Peoples R China 3.Univ Sci & Technol Beijing, Beijing, Peoples R China |
| 推荐引用方式 GB/T 7714 | Zhu, Jilong,Zhuo, Junbao,Wang, Shuhui. PIC: Domain generalization by path information constraint[J]. PATTERN RECOGNITION,2025,168:12. |
| APA | Zhu, Jilong,Zhuo, Junbao,&Wang, Shuhui.(2025).PIC: Domain generalization by path information constraint.PATTERN RECOGNITION,168,12. |
| MLA | Zhu, Jilong,et al."PIC: Domain generalization by path information constraint".PATTERN RECOGNITION 168(2025):12. |
| 条目包含的文件 | 条目无相关文件。 | |||||
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。
修改评论