Anisotropic boundary-aware detection for cotton leaf diseases with boundary-decoupled regression and lightweight feature adaptation.

Abstract

Detecting cotton leaf diseases in open-field environments is challenging due to cluttered backgrounds, scale variation, and irregular lesion morphology. Conventional detectors rely on isotropic receptive fields and coupled box-regression losses, which limit their ability to localize elongated lesions with poorly defined boundaries. We present an anisotropic boundary-aware detection framework that propagates high-frequency boundary information across four successive pipeline stages. In the backbone, an Anisotropic Morphological Contrast Aggregation module (AMCA) enhances direction-aware representation and lesion-background contrast via re-parameterizable strip convolutions and high-frequency residual extraction. A Dynamic Semantic Boundary Transfer mechanism (DSBT) then captures boundary priors from shallow layers before they are lost to downsampling and injects them into the neck. A Morphological-Spectral Synergistic Feature Pyramid Network (MFS-FPN) preserves these cues during multi-scale fusion through spatial-domain operations compatible with edge hardware. Finally, an Anisotropic Boundary-Decoupled IoU loss (ABD-IoU) independently penalizes each of the four box boundaries and sustains optimization signals in high-IoU regimes via a logarithmic modulation factor. On the self-constructed Complex Cotton Leaf Disease dataset (CCLD; 6,856 images, 6 classes), the method achieves 78.50% mAP@50 and 65.00% mAP@50:95, improving the YOLOv11n baseline by 4.80% and 2.70% with only 2.73 M parameters at 202 FPS. Cross-domain evaluations on PlantDoc and RWD confirm consistent improvements. The framework runs in real time on NVIDIA Jetson edge platforms with INT8 quantization.

Plant phenotyping relevance

綿花葉の病斑・病害状態を画像から検出する手法を開発し、複数データセットとベースラインで性能検証しているため、植物表現型取得が中心である。

Code and data availability

The authors explicitly state that their source code, trained models, and implementation details are publicly available, and the data availability statement points to the same repository, which hosts the self-constructed CCLD cotton leaf disease dataset (6,856 images, 6 classes) used for the paper's phenotyping/disease-