Crop leaf disease detection with additive gated convolution and hierarchical attention fusion.

Abstract

Crop leaf disease detection plays a crucial role in ensuring healthy crop growth and improving food security. Disease features are often small and have blurry edges, while background interference is strong, making precise detection a significant challenge. Although YOLO-based methods perform well in object detection, they still struggle to effectively handle the extraction of lesion details and background noise interference when applied to crop leaf disease detection. To address these challenges, this study introduces an innovative crop leaf disease detection approach built upon the YOLO, named AG-HAF. This method proposed two modules, the additive gated convolutional unit (AGCU), which introduces a gating mechanism to dynamically adjust the importance of features, enhancing the detection of small and blurry lesions, improving nonlinear feature modeling, and suppressing irrelevant background interference. Hierarchical Attention Fusion Module (HACFM), which utilizes a hierarchical attention mechanism to optimize the fusion of multi-scale features, enhancing the representation and semantic information of disease regions, and further improving the model's adaptability to complex backgrounds. Ablation and comparative experiments show that AG-HAF outperforms existing methods across various metrics, particularly excelling in disease detecting in complex backgrounds and small lesions, demonstrating significant potential for practical applications.

Plant phenotyping relevance

植物葉の病斑を画像から検出・推定する手法を新規開発し、比較実験とアブレーションで性能検証しているため、植物病害状態の画像ベース表現型計測が中心である。

Code and data availability

The paper uses two public leaf-disease image datasets, but both are cited third-party resources: the rice disease dataset is from Xiao et al. (2024, Kaggle DOI 10.34740/KAGGLE/DS/6336337) and the tomato leaves dataset is hosted on Roboflow (URL not among allowed_urls). No authors' own dataset, code, scripts, or trained

No evidence-backed public reproduction asset is currently recorded.