Domain Adaptation Across Geographic Regions Through Region-Specific Feature Learning and Distribution Matching
Deep neural networks achieve high accuracy in image recognition. However, they struggle with domain shifts, particularly from geographic variations. Although conventional domain adaptation methods address generalization across domains, such as real-to-clipart transfer, they do not address geographic...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11115023/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Deep neural networks achieve high accuracy in image recognition. However, they struggle with domain shifts, particularly from geographic variations. Although conventional domain adaptation methods address generalization across domains, such as real-to-clipart transfer, they do not address geographic variations for two key reasons: first, they assume uniform domain shifts across all classes, whereas geographic variations affect object categories non-uniformly, and second, they align global feature distributions without accounting for the selective influence of region-specific characteristics on different semantic categories. This challenge is compounded by the fact that most real-world datasets comprise images captured primarily in Western regions, which complicates generalization to Asian or other underrepresented regions. This study addresses geographic domain shift by incorporating both design and context shifts through a two-part approach: 1) a mechanism for learning region-specific features across object classes and 2)a distribution-matching technique that aligns features across regions while preserving class discrimination. In contrast to existing methods that eliminate all domain differences, our approach effectively distinguishes between universal object features and region-specific variations. We also introduce a novel feature alignment strategy at the superclass level that leverages hierarchical relationships between classes, enabling more effective knowledge transfer across regions. Experiments on the GeoNet benchmark show that the proposed framework outperforms existing methods, with improvements of up to 8.84 percentage points in top-1 accuracy and 17.26 percentage points in top-5 accuracy in cross-region tasks. Visualizations further confirm that the model attends to relevant image regions while incorporating appropriate contextual information for accurate classification when necessary. The reported implementation is available at <uri>https://github.com/meruemon/GDA</uri> |
|---|---|
| ISSN: | 2169-3536 |