Wang Xuewei, Liu Jun, Chen Qian
Shandong Provincial University Laboratory for Protected Horticulture, Weifang University of Science and Technology, Weifang, China.
School of Computer, Sichuan Technology and Business University, Chengdu, China.
Plant Methods. 2025 May 26;21(1):70. doi: 10.1186/s13007-025-01387-4.
Despite the significant progress in deep learning-based object detection, existing models struggle to perform optimally in complex agricultural environments. To address these challenges, this study introduces YOLO-Pepper, an enhanced model designed specifically for greenhouse pepper disease and pest detection, overcoming three key obstacles: small target recognition, multi-scale feature extraction under occlusion, and real-time processing demands. Built upon YOLOv10n, YOLO-Pepper incorporates four major innovations: (1) an Adaptive Multi-Scale Feature Extraction (AMSFE) module that improves feature capture through multi-branch convolutions; (2) a Dynamic Feature Pyramid Network (DFPN) enabling context-aware feature fusion; (3) a specialized Small Detection Head (SDH) tailored for minute targets; and (4) an Inner-CIoU loss function that enhances localization accuracy by 18% compared to standard CIoU. Evaluated on a diverse dataset of 8046 annotated images, YOLO-Pepper achieves state-of-the-art performance, with 94.26% mAP@0.5 at 115.26 FPS, marking an 11.88 percentage point improvement over YOLOv10n (82.38% mAP@0.5) while maintaining a lightweight structure (2.51 M parameters, 5.15 MB model size) optimized for edge deployment. Comparative experiments highlight YOLO-Pepper's superiority over nine benchmark models, particularly in detecting small and occluded targets. By addressing computational inefficiencies and refining small object detection capabilities, YOLO-Pepper provides robust technical support for intelligent agricultural monitoring systems, making it a highly effective tool for early disease detection and integrated pest management in commercial greenhouse operations.
尽管基于深度学习的目标检测取得了显著进展，但现有模型在复杂的农业环境中难以实现最佳性能。为应对这些挑战，本研究引入了YOLO - Pepper，这是一种专门为温室辣椒病虫害检测设计的增强模型，克服了三个关键障碍：小目标识别、遮挡情况下的多尺度特征提取以及实时处理需求。YOLO - Pepper基于YOLOv10n构建，包含四项主要创新：（1）自适应多尺度特征提取（AMSFE）模块，通过多分支卷积改进特征捕获；（2）动态特征金字塔网络（DFPN），实现上下文感知特征融合；（3）专为微小目标定制的专用小检测头（SDH）；（4）内部CIoU损失函数，与标准CIoU相比，定位精度提高了18%。在包含8046张标注图像的多样化数据集上进行评估，YOLO - Pepper达到了当前最优性能，在115.26 FPS下mAP@0.5为94.26%，比YOLOv10n（82.38% mAP@0.5）提高了11.88个百分点，同时保持了针对边缘部署优化的轻量级结构（251万个参数，5.15 MB模型大小）。对比实验突出了YOLO - Pepper相对于九个基准模型的优越性，特别是在检测小目标和被遮挡目标方面。通过解决计算效率低下问题并提升小目标检测能力，YOLO - Pepper为智能农业监测系统提供了强大的技术支持，使其成为商业温室运营中早期病害检测和综合虫害管理的高效工具。
