基于PDPI‑YOLO的辣椒病虫害识别算法

Pepper Disease and Pest Identification Algorithm Based on PDPI‑YOLO

摘要:

针对田间复杂环境下辣椒病虫害目标密集、尺度差异大以及现有检测模型部署难的问题，提出一种基于改进YOLO 11n的高效检测模型PDPI?YOLO。以YOLO 11n为基础，采用MobileNetV4重构主干，强化对病虫害的多尺度特征提取;引入SPPF_LSKA模块融合大核注意力，增强全局上下文信息以提升对小尺度害虫的捕捉能力;利用Dysample动态上采样优化特征融合，提高对形态多变目标的定位精度;集成EfficientHead检测头，在保障速度的同时增强对密集和跨尺度目标的检测性能。对Kaggle上公开的6类常见辣椒病虫害公开数据集开展模型训练和检测试验。结果表明，改进后模型精确率、召回率、mAP50和mAP50-95分别为80.8%、75.7%、83.1%和43.5%，相比原模型分别提高5.6、1.7、3.8、1.2个百分点，将改进后模型部署至可视化用户界面中进行部署检测，单幅图像推理时间在0.1 s以内，满足实时检测的要求。结合自建的2种东山光皮辣椒病虫害数据集进行检测试验，单幅图像推理时间在0.2 s以内，表明模型具备一定的实时性、泛化性和鲁棒性。该模型为辣椒病虫害自动防控系统的构建及移动端智能植保装备的研发提供了技术支撑。

Abstract:

Accurate identification of pepper diseases and pests in complex field environments remains challenging due to dense target distributions, significant scale variations, and practical deployment constraints of existing detection models. To address these issues, PDPI?YOLO, a efficient detection model was proposed based on an enhanced YOLO 11n architecture. Firstly, the model reconstructed the feature extraction structure using MobileNetV4 as the backbone, leveraging its efficient inverted residual blocks and integrated attention mechanisms to enhance multi?scale feature representation. Secondly, the original SPPF module in the backbone was replaced by the SPPF_LSKA feature enhancement module. By integrating the large kernel attention (LSKA) mechanism, this module augmented global contextual information at the end of the backbone, providing superior top?level semantic inputs for the feature pyramid and thereby optimizing the efficacy of feature fusion within the pyramid structure. Furthermore, a dynamic upsampling operator (Dysample) was incorporated into the feature fusion process to adaptively enhance multi?level feature integration and localization accuracy. Finally, the EfficientHead optimized the detection head structure, ensuring high precision while handling dense and scale?varying targets. Experiments on a public Kaggle dataset of six common pepper diseases and pests showed that PDPI?YOLO achieved a precision of 80.8%, recall of 75.7%, mAP50 of 83.1%, and mAP50-95 of 43.5%, outperforming the baseline by 5.6, 1.7, 3.8, and 1.2 percentage points, respectively. When deployed in a visual interface, the model achieved inference times under 0.1 s per image, meeting real?time requirements. Validation on a self?built dataset of Dongshan smooth?skinned pepper further confirmed its generalization capability and robustness, with inference times below 0.2 s per image. This model can offer practical technical support for automated disease and pest monitoring systems and intelligent plant protection equipment in edge environments.