| Title |
Domain Adaptive Plant-disease Segmentation via Color-driven Feature Distillation |
| Authors |
장소연(So-Yeon Jang) ; 류제호(Je-Ho Ryu) ; 김종옥(Jong-Ok Kim) |
| DOI |
https://doi.org/10.5573/ieie.2025.62.4.65 |
| Keywords |
Semantic segmentation; Domain adaptation; Color farthest point sampling; Correspondence loss; Plant-diease detection |
| Abstract |
As the demand for monitoring systems to enhance crop productivity grows, research on disease detection using UAV and close-range images has become increasingly active. However, existing studies often focus on models designed for specific scales or rely heavily on labeled data, limiting their applicability across diverse conditions. To address these limitations, this paper presents a domain-adaptive plant-disease segmentation model that integrates both UAV and close-range images, allowing for application across multiple scales. The proposed approach introduces a correspondence loss based on color similarity. This allows the model to effectively learn from unlabeled data by capturing feature and color information of similar classes, thereby improving class differentiation. Additionally, to address class imbalance during training, Color FPS (Color Farthest Point Sampling) is introduced to adjust the data balance across classes. This method allows the model to effectively capture the distinctions between healthy and diseased areas, thereby achieving robust performance under various imaging conditions. Experimental results on datasets collected from real-world agricultural settings demonstrate that the proposed method outperforms existing methods, highlighting its potential as a versatile model applicable in diverse environments. |