中国农业科学2025，Vol.58Issue(18)：3598-3615,18.DOI:10.3864/j.issn.0578-1752.2025.18.003
基于改进YOLOv8s的小麦苗期叶尖检测方法
Detection Method of Leaf Tip in Wheat Seedling Stage Based on Improved YOLOv8s
摘要
Abstract
[Objective]In precision agriculture,the detection of crop seedlings can be interfered with by factors such as soil weeds,occlusion between seedling leaves,and multi-scale datasets.Based on the object detection algorithm,this paper improved the YOLOv8s algorithm and designed the wheat leaf tip detection model YOLO-Wheat to solve problems,such as leaf occlusion of wheat seedlings in the field,interference from soil weeds,and multi-view data with multiple scales,thereby enhancing the accuracy of wheat seedling leaf detection and providing a theoretical basis for wheat seedling detection at the seedling stage in precision agriculture.[Method]Close-up and distant images of wheat seedlings were collected respectively through mobile phone cameras and on-board RGB cameras during the emergence period to construct a crop image dataset.In the network model,a pyramid structure of multi-scale feature fusion(high-level screening-feature fusion pyramid,HS-FPN)was adopted.This structure used high-level features as weights,filters low-level feature information through the channel attention module,and combined the screened features with the high-level features.Enhancing the feature expression ability of the model could effectively solve the problem of multi-scale data.Integrate the efficient local attention(ELA)local attention mechanism in the network model was used to enable the model to focus on the leaf tip information of wheat and to suppress the interference of soil background factors of weeds.Meanwhile,the loss function of YOLOv8s(complete IoULoss,CIoULoss)was optimized,and the inner-Iou Loss auxiliary bounding box loss function was introduced to enhance the network's attention to small targets and to improve the positioning accuracy of wheat leaf tips.In terms of training strategies,transfer learning was employed.The model was pre-trained using close-up images of wheat leaf tips,and then the parameters of the model were updated and optimized using distant images.[Result]The YOLO-Wheat model was compared with five object detection models,namely Faster-RCNN,YOLOv5s,YOLOv7,YOLOv8s,and YOLOv9s.The YOLO-Wheat model was the best in wheat leaf tip detection,with a recognition accuracy rate of 92.7%and a recall rate of 85.1%,respectively.The mean Average Precision(mAP)values were 82.9%.Compared with the Faster-RCNN,YOLOv5s,YOLOv7,YOLOv8s and YOLOv9s models,the recognition accuracy mAP values of YOLO-Wheat have increased by 17.1%,13.6%,11.0%,8.7%and 3.8%respectively;the recall rates increased by 13.1%,6.7%,4.5%,1.8%and 1.3%,respectively.Compared with the Faw-RCNN,YOLOv5s,YOLOv7,YOLOv8s and YOLOv9s models,the mAP values of YOLO-Wheat have increased by 16.2%,9.8%,5.0%,5.9%and 0.7%,respectively.[Conclusion]This method could effectively solve the problem of multi-scale data,achieve precise detection of small targets at the leaf tips of wheat seedlings in complex field environments using unmanned aerial vehicle(UAV)images,and provide technical support and theoretical reference for intelligent leaf counting of wheat seedlings in complex fields.关键词
小麦/YOLOv8s/损失函数/迁移学习/无人机Key words
wheat/YOLOv8s/loss function/transfer learning/drone引用本文复制引用
何豪旭,高祥,饶元,张子睿,吴巩,侯依廷,何烨,厉心怡..基于改进YOLOv8s的小麦苗期叶尖检测方法[J].中国农业科学,2025,58(18):3598-3615,18.基金项目
国家自然科学基金(32371993)、安徽省重点研究与开发计划(202204c06020026&2023n06020057)、安徽省高校自然科学研究重大项目(2022AH040125&2023AH040135) （32371993）
