基于轻量化改进型YOLOv5s的玉米病虫害检测方法

A corn disease and pest detection method based on lightweight improved YOLOv5s

针对复杂环境下目前现有的玉米病虫害检测方法的精度不理想、模型复杂、难以在移动端部署等问题,本研究提出了基于轻量化改进型YOLOv5s的玉米病虫害检测方法。首先,采用轻量级网络GhostNet替换原始YOLOv5s模型中特征提取网络和特征融合网络的卷积层,降低模型的计算量和参数量,提高运行速度,以满足移动端的部署要求;其次,为弥补GhostNet所带来的检测精度下降缺陷,在模型的主干特征提取网络中引入注意力机制,更加全面地评估特征权值,以增强玉米病虫害的特征,减弱无关信息的干扰,提升检测性能;最后,将模型的损失函数由CIOU替换为EIOU,以增强模型对目标的精确定位能力,从而提升模型的收敛速度和回归精度。试验结果表明,改进模型相比原始YOLOv5s模型在对供试玉米病虫害检测中,P、R和mAP分别提高了1.9个百分点、2.2个百分点和2.0个百分点,分别达到了94.6%、80.2%和88.8%;在保持较高检测精度的同时,模型的计算量、参数量和模型大小分别减少了50.6%、52.9%和50.4%,解决了检测模型在移动端的部署问题。

Aiming at the problems of unsatisfactory detection accuracy,complex model and difficult deployment on mobile terminals in the existing maize disease and pest detection methods in complex environments,this study proposed a maize disease and pest detection method based on lightweight improved YOLOv5s.Firstly,the lightweight network GhostNet was used to replace the convolutional layer in the feature extraction network and feature fusion network in the original YOLOv5s model,which reduced the calculation and parameter amount of the model and improved the running speed to meet the deployment requirements of the mobile terminal.Secondly,in order to compensate for the problem of detection accuracy degradation caused by GhostNet,the normalization-based attention module(NAM)was introduced into the backbone feature extraction network of the model to evaluate the feature weights more comprehensively,so as to enhance the characteristics of corn diseases and pests,weaken the interference of irrelevant information,and improve the detection performance.Finally,the loss function of the model was replaced by EIOU from CIOU to enhance the model’s ability to accurately locate the target,so as to improve the convergence speed and regression accuracy of the model.The experimental results showed that compared with the original YOLOv5s model,the P,R and mAP of the final improved model increased by 1.9 percentage points,2.2 percentage points and 2.0 percentage points,respectively,reaching 94.6%,80.2%and 88.8%.While maintaining high detection accuracy,the calculation amount,parameter amount and capacity of the model were reduced by 50.6%,52.9%and 50.4%,which solved the deployment problem of the detection model on the mobile terminal.