基于UD-DETR轻量化网络小目标异物检测算法  

UD-DETRA Lightweight Network Algorithm for the Detection of Small Foreign Objects

作　　者：林永洪 郑建明[1] 李照 Lin Yonghong;Zheng Jianming;Li Zhao(School of Information and Control Engineering,Jilin Institute of Chemical Technology,Jilin 132022,China;Jilin Urban and Rural Planning Research Institute,Jilin 132000,China)

机构地区：[1]吉林化工学院信息与控制工程学院,吉林132022 [2]吉林市城乡规划研究院,吉林132000

出　　处：《信息化研究》2025年第1期71-78,共8页INFORMATIZATION RESEARCH

摘　　要：RT-DETR作为DETR系列中针对实时场景进行优化的模型,同时存在模型较大、对多尺度的小目标异物存在误检、漏检的问题。为了降低模型的参数量、计算量,同时提高多尺度特征提取能力,本文提出一种基于RT-DETR改进的UD-DETR传送带异物检测方法,SFDConv利用不同类型的卷积分散计算成本和捕捉更多样化的特征信息,用于改进主干网络残差块中的3×3卷积;提出的UDFPN能够降低路径聚合的计算成本和增强提取小目标在空间和深度方向上的局部特征的能力,用于改进高效混合编码器。实验表明,本文所提出的UD-DETR的参数量为11 M,FLOPS为39.4 G,精确度为0.950,召回率为0.836,mAP@50为0.899,mAP@50-95为0.522。参数量和FLOPS相较于原版模型下降了44%和30%,性能指标均高于原版模型和YOLOv8、YOLOv9等主流检测模型。RT-DETR,as a model optimized for real-time scenarios in the DETR family,also suffers from a large model size,misdetection and omission of small-target foreign objects at multiple scales.In order to re-duce the number of parameters and computation of the model,as well as to improve the multi-scale feature ex-traction capability,this paper proposes a conveyor belt foreign object detection method based on RT-DETR im-proved UD-DETR.The proposed SFDConv utilizes different types of convolutions to spread the computational cost and capture more diverse feature information for improving 3×3 convolutions in residual blocks of the backbone network;The proposed UDFPN reduces the computational cost of path aggregation and enhances the ability to extract local features of small targets in both spatial and depth directions for improving efficient hybrid encoders.The experiments show that the proposed UD-DETR have 11m parameters,39.4 FLOPS,0.95 preci-sion,0.836 recall,0.899 mAP@50,and 0.522 [e-mail] number of parameters and the FLOPS drop by 44%and 30%compared to the original model,and the performance measures are higher than those of the original model,YOLOv8,YOLOv9 and other mainstream detection models.

关 键 词：RT-DETR UD-DETR 传送带异物检测 SFDConv UDFPN 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]