基于改进YOLO v8s的水稻种植机械作业质量检测  

Rice Planting Machinery Operation Quality Detection Based on Improved YOLO v8s

作　　者：刘双喜[1,2] 张玮平 胡宪亮 王刘西航 宋占华 王金星[1,4] LIU Shuangxi;ZHANG Weiping;HU Xianliang;WANG Liuxihang;SONG Zhanhua;WANG Jinxing(College of Mechanical and Electronic Engineering,Shandong Agricultural University,Taian 271018,China;Shandong Engineering Research Center of Agricultural Equipment Intelligentization,Taian 271018,China;Jinan Xiangchen Technology Co.,Ltd.,Ji’nan 251400,China;Shandong Key Laboratory of Intelligent Production Technology and Equipment for Facility Horticulture,Taian 271018,China)

机构地区：[1]山东农业大学机械与电子工程学院,泰安271018 [2]农业装备智能化山东省工程研究中心,泰安271018 [3]济南祥辰科技有限公司,济南251400 [4]山东省设施园艺智慧生产技术装备重点实验室(筹),泰安271018

出　　处：《农业机械学报》2024年第S1期61-70,共10页Transactions of the Chinese Society for Agricultural Machinery

基　　金：山东省现代农业产业技术体系水稻农业机械岗位专家项目(SDAIT-17-08)

摘　　要：稻田中秧苗与稻种规范化精准识别检测是实现水稻种植机械作业质量检测的前提,为解决水稻种植图像识别研究过程中存在稻田背景复杂、机械运行速度快、形态特征难以提取等造成识别准确率较低的问题,提出一种基于改进YOLO v8s的轻量化质量检测方法。首先,通过由井关PZ60型水稻插秧机的研制而成的稻田种植质量检测装置,搭建作业质量检测图像采集平台,拍摄获得作业质量的图像构成ImageSets数据集,根据国家相关标准制定质量检测评价指标。随后通过引入轻量化GhostNet模块,减少网络模型的运行参数量;同时为了提升卷积神经网络检测性能,将CPCA注意力模块引入到检测算法中,有效地增强对水稻作业质量的特征提取,抑制稻田复杂的背景信息,准确获得作业图像的关键特征,对秧苗与稻种这种数量多、体积小的目标的检测效果有较为明显的提升;其次,将YOLO v8s模型中的CIoU损失函数替换为EIoU损失函数,使模型具有更快更好的收敛速度与定位效果,实现作业质量的精确识别。试验结果表明,改进后的YOLO v8s模型在测试集上的平均精度均值为92.41%,精确率为92.11%,召回率为92.04%;与YOLO v5s、YOLO v7、YOLO v8s、Faster R-CNN网络模型相比,平均精度均值分别提高7.91、7.71、4.28、1.03个百分点。改进后模型检测速度与内存占用量分别为88 f/s、19.2 MB,与YOLO v8s模型相比分别减少12.8%、10.7%,经种植环境测试能够检测出作业质量是否合格,能够实现质量检测的作用。改进YOLO v8s网络模型对稻田作业质量检测具有快速准确的识别能力,具有较好的鲁棒性,在水稻种植质量检测方面有显著成效,可为水稻种植机械化质量检测提供新的检测方法。The standardized and precise identification and detection of seedlings and seeds in rice fields is a prerequisite for achieving the quality detection of mechanical rice planting operations.To address the issues of complex rice field backgrounds,high machinery operation speeds,and difficulty in extracting morphological features during the research on rice planting image recognition,which resulted in low recognition accuracy rates,a lightweight quality detection method based on the improved YOLO v8s was proposed.Firstly,an image acquisition platform for operation quality detection was established through a rice planting quality detection device developed from the Inaka PZ60 type rice transplanter.Images of operation quality were captured to form the ImageSets dataset,and quality detection evaluation indicators were formulated in accordance with relevant national standards.Then by introducing the lightweight GhostNet module,the operational parameters of the network model were reduced.Simultaneously,to enhance the detection performance of the convolutional neural network,the CPCA attention module was incorporated into the detection algorithm,effectively strengthening the feature extraction for the quality of rice planting operations,suppressing the complex background information of the rice field,accurately obtaining the key features of the operation images,and significantly improving the detection effect of numerous small targets such as seedlings and seeds.Secondly,the CIoU loss function in the YOLO v8s model was replaced with the EIoU loss function,enabling the model to have a fast and good convergence speed and localization effect,and achieving precise identification of operation quality.The experimental results indicated that when evaluated using the average precision as the main indicator,the average precision of the improved YOLO v8s model on the test set was 92.41%,with an accuracy of 92.11%,a recall of 92.04%,and an mAP improvement of 7.91,7.71,4.28,and 1.03 percentage points,respectively,compared with the Y

关 键 词：水稻种植质量检测 机械作业 YOLO v8 漂秧漏秧 图像识别 

分 类 号：S223[农业科学—农业机械化工程] TP391.41[农业科学—农业工程]