基于改进YOLO v8n的辣椒穴盘育苗播种质量轻量级检测方法  

Lightweight Detection Method for Seeding Quality of Chili Seedling Trays Based on Improved YOLO v8n

作　　者：孔德航 刘云强 崔巍[1,2] 吴海华 张学东[1,2] 宁义超 KONG Dehang;LIU Yunqiang;CUI Wei;WU Haihua;ZHANG Xuedong;NING Yichao(Chinese Academy of Agricultural Mechanization Sciences Group Co.,Ltd.,Beijing 100083,China;State Key Laboratory of Agricultural Equipment Technology,Beijing 100083,China)

机构地区：[1]中国农业机械化科学研究院集团有限公司,北京100083 [2]农业装备技术全国重点实验室,北京100083

出　　处：《农业机械学报》2025年第2期381-392,共12页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家重点研发计划项目(2023YFD2001200)。

摘　　要：针对辣椒穴盘育苗播种质量实时、准确检测难和边缘设备算力有限等问题,基于YOLO v8n设计了一种轻量级检测算法YOLO v8n-SCS(YOLO v8n improved with StarNet,CAM and SCConv)。采用“单格训练+整盘检测”的技术策略以降低训练成本,提高训练效率。采用StarNet轻量级网络和上下文增强模块(Context augmentation module,CAM)作为主干网络,在降低模型复杂程度同时,实现深层特征多感受野信息融合;采用空间与通道重建卷积(Spatial and channel reconstruction convolution,SCConv)优化中间层C2f(Cross stage partial network fusion)模块的瓶颈结构,增强模块特征提取能力和提高模型推理速度;融合P2检测层,将基线3个检测头减至1个,增强模型对小目标的检测性能。结果表明,YOLO v8n-SCS模型参数量为1.2×10^(6)、内存占用量为2.7 MB、浮点数运算量为7.6×10^(9),在穴盘单格数据集上,其mAP_(50)为98.3%、mAP_(50-95)为83.8%、帧率为112 f/s,相比基线模型YOLO v8n,参数量降低62.5%、mAP_(50)提升2.5个百分点、mAP_(50-95)提升2.1个百分点、浮点数运算量降低14.6%、帧率提升23.1%;在整盘检测任务中,其检测帧率为21 f/s,检测准确率为98.2%,相比基线模型检测帧率提升8.2%、准确率提升1.1个百分点,对于播种速度800盘/h以内的72穴育苗盘和600盘/h以内的128穴育苗盘,其平均检测准确率大于96%,且单粒率、重播率和漏播率检测误差小于1.8%。本文模型在性能和计算成本之间取得了很好的平衡,降低了部署边缘设备计算要求,满足辣椒穴盘育苗播种质量在线检测需求,为育苗播种生产线智能化升级提供了技术支持。Aiming to address the challenges of real-time and accurate detection of the seeding quality of chili seedling trays, considering the computing power limitations of edge devices, a lightweight detection algorithm YOLO v8n-SCS(YOLO v8n improved with StarNet, CAM, and SCConv) was proposed based on YOLO v8n. Meanwhile, the technical strategy of “single-cell training + whole-tray detection” was adopted to reduce training costs and improve training efficiency. Firstly, the StarNet lightweight network and the CAM(Context augmentation module) were used as the backbone network to achieve multi-receptive field information fusion of deep features while reducing the complexity of the model. Secondly, the spatial and channel reconstruction convolution(SCConv)was employed to optimize the bottleneck structure of the intermediate layer cross stage partial network fusion(C2f) module to enhance the feature extraction ability of the module and improve the model inference speed. Finally, the P2 detection layer was fused and the detection heads were reduced to one to enhance the model's detection performance for small targets. The results showed that the YOLO v8n-SCS model had a parameter quantity of 1.2×10^(6), a memory occupation of 2.7 MB, and a computation amount of 7.6×10^(9) FLOPs. On the single-cell dataset of the seedling trays, its mAP_(50) was 98.3%, mAP_(50-95) was 83.8%, and the frame rate was 112 f/s. Compared with the baseline model YOLO v8n, the parameter quantity was reduced by 62.5%, mAP_(50) was increased by 2.5 percentage points, mAP_(50-95) was increased by 2.1 percentage points, the floating-point operations were reduced by 1.3×10^(9), and the frame rate was increased by 23.1%. In the whole-tray detection task, its detection frame rate was 21 f/s and the detection accuracy rate was 98.2%. Compared with the baseline model, the detection frame rate was increased by 8.2% and the accuracy rate was increased by 1.1 percentage points. For 72-cell seedling trays with a seeding speed within 800 trays/h and 128-cel

关 键 词：辣椒种子 穴盘育苗 播种质量检测 改进YOLO v8n 轻量级模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] S223[自动化与计算机技术—计算机科学与技术]