基于YOLO的面向中药材病虫害检测的轻量化自适应注意力模型

Lightweight Self-Adaptive Attention Model for Traditional Chinese Medicinal Herb Pest/Disease Detection Based on YOLO

摘要: 中药材病虫害检测面临小目标高密度分布与复杂背景干扰的双重挑战。本文提出一种轻量化自适应检测模型Agri-YOLO，通过多尺度自适应注意力(MAA)模块动态融合局部纹理与全局语义特征，并设计Ghost-ECA轻量化主干网络，在降低模型复杂度的同时提升小目标检测能力。实验表明，在自建数据集上，Agri-YOLO的mAP@0.5达到89.4% (较YOLOv8n提升6.2%)，小目标漏检率从35.7%显著降至13.1%，参数量仅1.5 M (减少52%)，计算量降至3.4 GFLOPs (降低61%)。该模型为复杂农业场景下的实时病虫害检测提供了高效解决方案。

Abstract: Pest detection in traditional Chinese medicinal herbs faces dual challenges of small objects with high-density distribution and complex background interference. This paper proposes Agri-YOLO, a lightweight adaptive detection model. The model dynamically integrates local texture and global semantic features through a Multi-Scale Adaptive Attention (MAA) module, and employs a Ghost-ECA lightweight backbone network to reduce model complexity while enhancing small object detection capability. Experimental results on our self-built dataset show that Agri-YOLO achieves a mAP@0.5 of 89.4% (6.2% higher than YOLOv8n), reduces the small object miss rate from 35.7% to 13.1%, and maintains only 1.5 M parameters (52% reduction) with 3.4 GFLOPs computational complexity (61% reduction). This model provides an efficient solution for real-time pest detection in complex agricultural scenarios.