摘要
针对传统油菜虫害检测方法存在的作业效率低、过度依赖主观经验、准确性难以保证等问题,基于YOLOv8模型提出了一种协同知识蒸馏轻量化检测方法(Feature and Logit Knowledge Distill,FLKD)。该方法以YOLOv8s为教师模型、以YOLOv8n为学生模型,采用logit蒸馏与特征蒸馏融合的改进策略,首先通过掩码生成式蒸馏MGD重构教师模型的多尺度特征,利用轻量级3×3卷积模块实现剪枝后特征图的自适应映射;其次将学生模型特征输入教师检测头进行预测对齐,通过最小化输出分布差异传递高阶语义决策知识。对比实验结果表明,FLKD在自建油菜虫害数据集ACEFP中的mA P@0.5达到96.7%,且参数量从11.2M压缩至4.4M,相较其他方法具有明显的优势,为田间油菜虫害的实时检测提供了精度与效率均衡的轻量化解决方案。
A collaborative knowledge distillation lightweight detection method(Feature and Logit Knowledge Distillation,FLKD)based on the YOLOv8 model is proposed to address the problems of low operational efficiency,excessive reliance on subjective experience,and difficulty in ensuring accuracy in traditional rapeseed pest detection methods.This method uses YOLOv8s as the teacher model and YOLOv8n as the student model,and adopts an improved strategy of logit distillation and feature distillation fusion.Firstly,the multi-scale features of the teacher model are reconstructed through mask generative distillation MGD,and the adaptive mapping of the pruned feature map is achieved using a lightweight 3×3 convolution module.Secondly,the student model features are input into the teacher detection head for prediction alignment,and high-order semantic decision knowledge is transmitted by minimizing the output distribution differences.The comparative experimental results show that FLKD achieves a mAP@0.5 of 96.7%in the self built rapeseed pest dataset ACEFP,and the parameter size is compressed from 11.2M to 4.4M,which has significant advantages compared to other methods.It provides a lightweight solution that balances accuracy and efficiency for real-time detection of rapeseed pests in the field.
作者
罗庆庭
卓辉
LUO Qingting;ZHUO Hui(College of Information and Intelligence,Hunan Agricultural University,Changsha,Hunan 410128,China)
出处
《农业工程与装备》
2024年第6期18-22,共5页
AGRICULTURAL ENGINEERING AND EQUIPMENT
