计算机工程与应用2026，Vol.62Issue(1)：112-123,12.DOI:10.3778/j.issn.1002-8331.2503-0277
基于改进RT-DETR的锻件表面缺陷检测算法
Surface Defect Detection Algorithm for Forgings Based on RT-DETR
摘要
Abstract
The forging surface defects are harmful with low detection efficiency.To address the existing problems of forg-ing surface defects detection,this paper proposes an algorithm based on improved RT-DETR.First,it collects magnetic particle inspection images as datasets in the vehicle steering knuckle manufacturing workshop of Hubei Sanhuan Forging Co.Then,a lightweight cross-stage heat conduction module is proposed,which introduces a simulated heat diffusion pro-cess into the frequency-domain modeling mechanism to achieve global perception and suppress high-frequency noise;meanwhile,it introduces context guide feature pyramid network to realize multi-scale feature fusion through dynamic channel alignment and spatial attention guidance,thereby enhancing semantic consistency and contextual integration of targets;finally,it uses dynamic position bias(DPB)module to enhance the extracting ability of cross-scale features.The experimental results on the forging surface crack dataset show that the mAP value reaches 87.9%,and the parameters and FLOPs are reduced by 20.7%and 9.3%,which is better than other mainstream algorithms.On the NEU-DET dataset,the improved RT-DETR model improves 1.2 percentage points in mAP compared to the benchmark model,which proves that the algorithm is generalizable.In conclusion,the algorithm has improved accuracy and reduced model complexity,it is suitable for deployment and application in real manufacturing situations.关键词
实时检测转换器(RT-DETR)/缺陷检测/特征提取/锻件/动态位置偏置模块Key words
real-time detection transformer(RT-DETR)/defect detection/feature extraction/forgings/dynamic positional bias module分类
信息技术与安全科学引用本文复制引用
张国文,张上,张岳,李琼,张军..基于改进RT-DETR的锻件表面缺陷检测算法[J].计算机工程与应用,2026,62(1):112-123,12.基金项目
湖北省数字经济试点示范建设专项(2312-420625-04-02-996363) （2312-420625-04-02-996363）
湖北省国家级大学生创新创业训练计划(S202311075047) （S202311075047）
国家级大学生创新创业训练计划(202111075012,202011075013). （202111075012,202011075013）
