Answers are grounded in this paper and cite their sources.
Corn Southern Leaf Blight (CSLB) causes substantial yield losses and quality deterioration in corn, and breeding resistant varieties is one of the most effective strategies for disease control. This breeding process relies on high-precision lesion segmentation to quantify disease severity at the pixel level. However, this task remains challenging due to the small and irregular lesion morphology, blurred boundaries, and low contrast against healthy leaf tissues, especially for the severe class imbalance inherent to the task. To address these challenges, this study proposes BPNet, a boundary-aware and prototype-constrained network for CSLB lesion segmentation. Specifically, we design two complementary modules to enhance the final decoded features: a mask-boundary head (MBH) to improve lesion contour delineation through joint mask-boundary prediction, and prototype regularization (PR) to enforce feature consistency and lesion-background separability in the embedding space. We further develop a region-adaptive KAN block (RAKB) that employs a soft-gated dual-expert structure with a spatial gating mechanism to capture complementary lesion and background feature representations. We also construct a CSLB image dataset to support breeding programs, providing expert-verified, pixel-level annotations that enable accurate lesion area ratio quantification. Experimental results demonstrate that BPNet outperforms existing segmentation methods in delineating fine lesion boundaries and detecting small lesions, while maintaining a compact model size and low computational cost, indicating its potential to support efficient lesion quantification under standardized imaging conditions. Our dataset and code are available at https://github.com/hpguo1982/BPNet .
No takes yet. Share an insight, caveat, or question.
Sun et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: