基于改进Detection Transformer的棉花幼苗与杂草检测模型研究  

Research on Cotton Seedling and Weed Detection Model Based on Improved Detection Transformer

作　　者：冯向萍[1] 杜晨 李永可[1] 张世豪 舒芹 赵昀杰 FENG Xiangping;DU Chen;LI Yongke;ZHANG Shihao;SHU Qin;ZHAO Yunjie(School of Computer Science and Engineering,Xinjiang Agricultural University,Urumqi 830052;School of Economics and Management,Xinjiang Agricultural University,Urumqi 830052)

机构地区：[1]新疆农业大学计算机与信息工程学院,乌鲁木齐830052 [2]新疆农业大学经济管理学院,乌鲁木齐830052

出　　处：《计算机与数字工程》2024年第7期2176-2182,共7页Computer & Digital Engineering

基　　金：科技部科技创新2030-“新一代人工智能”重大项目(编号:2022ZD0115805);新疆维吾尔自治区重大专项(编号:2022A02011-3)资助。

摘　　要：基于深度学习的目标检测技术在棉花幼苗与杂草检测领域已取得一定进展。论文提出了基于改进Detection Transformer的棉花幼苗与杂草检测模型,以提高杂草目标检测的准确率和效率。首先,引入了可变形注意力模块替代原始模型中的Transformer注意力模块,提高模型对特征图目标形变的处理能力。提出新的降噪训练机制,解决了二分图匹配不稳定问题。提出混合查询选择策略,提高解码器对目标类别和位置信息的利用效率。使用Swin Transformer作为网络主干,提高模型特征提取能力。通过对比原网络,论文提出的模型方法在训练过程中表现出更快的收敛速度,并且在准确率方面提高了6.7%。Significant progress has been made in the field of cotton seedling and weed detection using deep learning-based ob⁃ject detection techniques.This article proposes an improved Detection Transformer-based model for cotton seedling and weed detec⁃tion to improve the accuracy and efficiency of weed target detection.Firstly,a deformable attention module is introduced to replace the Transformer attention module in the original model,improving the model's ability to handle feature map object deformation.A new denoising training mechanism is proposed to address the unstable bipartite graph matching problem.A hybrid query selection strategy is proposed to improve the decoder's utilization efficiency of target category and position information.The Swin Transformer is used as the network backbone to enhance the model's feature extraction ability.By comparing with the original network,the pro⁃posed model demonstrates a faster convergence speed during training and improves accuracy by 6.7%.

关 键 词：目标检测 Detection Transformer 棉花幼苗 杂草检测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP312[自动化与计算机技术—控制科学与工程]