Chen Chen, Lu Xiaolei, He Lei, Xu Ruoxue, Yang Yi, Qiu Jing
College of Big Data (College of Information Engineering), Yunnan Agricultural University, Kunming, China.
The Key Laboratory for Crop Production and Smart Agricultural of Yunnan Province, Yunnan Agricultural University, Kunming, China.
Front Plant Sci. 2025 Apr 7;16:1523633. doi: 10.3389/fpls.2025.1523633. eCollection 2025.
The rapid and accurate identification of soybean diseases is critical for optimizing both yield and quality. Traditional image recognition techniques face notable limitations in terms of generalization and accuracy, particularly when tasked with identifying small-scale targets or distinguishing diseases with similar characteristics in large, heterogeneous, and complex environments. To address these challenges, this study proposes the YOLOv8-DML model for soybean leaf disease recognition. Building upon YOLOv8n, this model integrates a DWR module that replaces the high-level C2f module with C2f-DWR, enhancing feature extraction across varied receptive fields. Additionally, modifications to the neck structure incorporate a Multi-scale Enhanced Feature Pyramid (MEFP), which improves detection performance across targets of varying sizes by enabling effective multi-scale information fusion. A lightweight detection head (LSCD) is further introduced to facilitate multiscale feature interactions while reducing the overall model parameter count. Lastly, the WIoUv3 loss function is employed to place greater emphasis on small targets and moderate-quality samples, thereby enhancing detection precision. Experimental results demonstrate that YOLOv8-DML achieves a mAP50 of 96.9%, marking a 1.8% improvement over the original YOLOv8 algorithm, while also achieving an 18.6% reduction in parameters. Comparative analysis with other mainstream object detection models indicates that YOLOv8-DML delivers superior overall performance, highlighting its significant potential for effective soybean leaf disease identification.
快速准确地识别大豆病害对于优化产量和品质至关重要。传统的图像识别技术在泛化能力和准确性方面存在显著局限性，尤其是在识别小规模目标或在大型、异质且复杂的环境中区分具有相似特征的病害时。为应对这些挑战，本研究提出了用于大豆叶部病害识别的YOLOv8-DML模型。该模型基于YOLOv8n构建，集成了一个DWR模块，该模块用C2f-DWR取代了高级C2f模块，增强了跨不同感受野的特征提取。此外，对颈部结构的修改引入了多尺度增强特征金字塔（MEFP），通过实现有效的多尺度信息融合来提高对不同大小目标的检测性能。还引入了一个轻量级检测头（LSCD），以促进多尺度特征交互，同时减少整体模型参数数量。最后，采用WIoUv3损失函数来更加强调小目标和中等质量样本，从而提高检测精度。实验结果表明，YOLOv8-DML的mAP50达到96.9%，比原始YOLOv8算法提高了1.8%，同时参数减少了18.6%。与其他主流目标检测模型的对比分析表明，YOLOv8-DML具有卓越的整体性能，凸显了其在有效识别大豆叶部病害方面的巨大潜力。
