Color Guided Disease Segmentation for Plant Images

This paper proposes a network architecture designed for semantic disease segmentation on plant images for smart agriculture. To effectively segment challenging small disease regions, the proposed method incorporates hierarchical fusions with color information into the conventional encoder-decoder network. The color information within input images can help the network to easily find regions associated with a target class. We extract color information by color layer decomposition, and it works as guidance for semantic features during the fusion process. The hierarchical fusion of encoder features and color features is highly informative, highlighting each class and enhancing disease segmentation performance. The proposed method exceeds existing semantic segmentation methods and handles small areas associated with diseases more effectively.