摘要
精准检测果实的生长阶段对于提高农业机器人的作业效率至关重要。然而,传统的全监督目标检测模型依赖于大量的数据标注,实现成本过高。鉴于此,提出了一种利用少量标注数据训练的目标检测模型。包括:(1)采用YOLOv8构建教师-学生模型,负责生成伪标签与训练网络;(2)通过改良的伪标签分配方法,筛选出可靠、不确定与不可靠的伪标签,减少低质量伪标签对学生网络训练的负面影响;(3)针对这三种伪标签优化了训练损失函数,以更好地提取不同置信度的伪标签内的潜在信息。通过在石榴生长阶段数据集上进行实验验证,结果显示,在仅使用10%的标注数据的情况下,所提方法的平均精度达到了90.8%,比同样数据集下的YOLOv8网络提升了5.5个百分点,且该性能已接近于使用全标注数据训练的YOLOv8的92%。这一结果充分证明了提出的方法在标注资源有限的背景下,依然能够训练出高效且准确的目标检测模型。
Accurate detection of fruit growth stages is crucial for improving the operational efficiency of agricultural robots.However,traditional fully supervised object detection models rely on a large amount of data labeling,which is too costly to implement.In view of this,this study proposes a object detection model trained using a small amount of labeled data.It includes:(1)constructing a teacher-student model using YOLOv8,which is responsible for generating pseudo-labeling and training the network;(2)filtering out reliable,uncertain and unreliable pseudo-labels through an improved pseudo-label assignment method to reduce the negative impact of low-quality pseudo-labels on the training of student networks;(3)optimizing the training loss function for these three kinds of pseudo-labels to better extract potential information within the pseudo-labels with different confidence levels.Through experimental validation on the pomegranate growth stage dataset,the results show that the proposed method achieves an average accuracy of 90.8% with only 10% labeled data,which is 5.5 percentage points higher than the YOLOv8 network under the same dataset,and this performance is close to the 92% of the YOLOv8 trained with fully labeled data.This result fully demonstrates that the method proposed in this study can still train efficient and accurate object detection models in the context of limited labeling resources.
作者
张立才
张欣
陈孝玉龙
刘佳明
文兴甜
杨胜贤
ZHANG Licai;ZHANG Xin;CHEN Xiaoyulong;LIU Jiaming;WEN Xingtian;YANG Shengxian(College of Big Data and Information Engineering,Guizhou University,Guiyang 550025,China;College of Life Sciences,Guizhou University,Guiyang 550025,China)
出处
《计算机工程与应用》
北大核心
2025年第13期291-299,共9页
Computer Engineering and Applications
基金
国家重点研发计划重点专项(2021YFE0107700)
贵州省科技厅平台项目(黔科合平台人才-HZD[2022]001)。
关键词
半监督
伪标签
目标检测
师生模型
semi-supervised
pseudo-label
object detection
teacher-student model
作者简介
张立才(1993-),男,硕士研究生,研究方向为农业信息化与图像处理;通信作者:张欣(1976-),男,副教授,研究方向为智慧农业,E-mail:Xzhang1@gzu.edu.cn;陈孝玉龙(1988-),男,教授,研究方向为植物保护与生物学;刘佳明(1997-),男,硕士研究生,研究方向为图像处理;文兴甜(2000-),男,硕士研究生,研究方向为植物表型分析;杨胜贤(2001-),男,硕士研究生,研究方向为农业大数据与图像处理。
