基于轻量化YOLOv8-Rice的库尔勒香梨虫害监测研究  

Study on pest monitoring of Korla Fragrant Pear based on lightweight YOLOv8-Rice

作　　者：周鹏 曹冰玉 杨权鑫 Zhou Peng;Cao Bingyu;Yang Quanxin(College of Information Science and Engineering,Xinjiang University of Science and Technology,Korla,841000,Xinjiang,China)

机构地区：[1]新疆科技学院信息科学与工程学院,新疆库尔勒841000

出　　处：《新疆农机化》2025年第2期32-36,共5页Xinjiang Agricultural Mechanization

基　　金：新疆科技学院2024年产教融合与新商科发展研究中心第二批招标项目(2024-KYJD05);校级创新团队专项(2024-KYTD01)。

摘　　要：针对库尔勒香梨病虫害识别中背景复杂、计算量大以及模型参数量大,且难以在移动或嵌入式设备上部署等挑战,本研究在YOLOv8的基础上提出了一种改进的轻量化模型—YOLOv8-Rice。该模型采用ContextGuidedBlock结构替换了YOLOv8中C2f模块的Bottleneck结构,有效增强了模型对上下文信息的理解能力,并压缩了模型权重。其次使用深度可分离卷积替代YOLOv8中的标准卷积,显著降低了模型的参数量和计算量,通过重构检测头为轻量级共享卷积检测头,进一步降低了模型的参数量和计算量,并提升了模型对多尺度病虫害特征的定位和提取能力。试验结果表明,与YOLOv8相比YOLOv8-Rice在计算量和参数量方面分别减少了71.1%和61.5%,模型权重文件大小降至1.89 MB,仅为YOLOv8的32.4%,同时在平均精度上达到了94.0%,与其他模型相比有明显提升。Aiming at the challenges of complex background,large calculation amount and large parameters of the model and difficult to be deployed on mobile or embedded devices in the identification of pests and diseases of Korla Fragrant Pear,this study proposed an improved lightweight model(YOLOv8-Rice)based on YOLOv8.Firstly,this model used the Context Guided Block structure to replace the Bottlecheck structure of the C2f module in YOLOv8,which effectively enhanced the understanding ability of the model to the context information and compresses the model weight.Then,the standard convolution in YOLOv8 was replaced by the depth-separatable convolution,which significantly reduced the parameters and calculation of the model.Finally,by reconfiguring the detection head as a lightweight shared convolution detection head,the parameter and calculation amount of the model were further reduced,and the positioning and extraction capacity of the model for the characteristics of multi-scale diseases and insect pests was improved.The experimental results show that,compared with YOLOv8,the computational and parameter quantities of YOLOv8-Rice are reduced by 71.1%and 61.5%respectively,and that the model weight file size is reduced to 1.89 MB,only 32.4%of YOLOv8.Meanwhile,the average accuracy reaches 94.0%,which is significantly improved compared with other models.

关 键 词：病虫害检测 深度学习 库尔勒香梨 轻量化 

分 类 号：S24[农业科学—农业电气化与自动化] TP2[农业科学—农业工程]