PAM-DETR: 基于改进RT-DETR的医用手套小目标缺陷检测算法

doi:10.16088/j.issn.1001-6600.2025112602

摘要/Abstract

摘要： 针对医用手套表面缺陷检测中存在的微小目标识别精度低、背景干扰强等问题,本文提出一种基于Transformer架构的轻量化检测模型PAM-DETR(parameter-efficient adaptive multi-scale detection transformer)。首先,在RT-DETR模型基础上构建PRT-Block模块,引入重参数化结构与注意力机制,在剔除标准卷积计算冗余的同时显著增强小目标特征的表达能力;其次,设计自适应稀疏编码模块ASEM(adaptive sparse encoder module)替代原始AIFI结构,优化特征尺度的交互效率;再次,融合ASEM与全新设计的CSPOmniKernel结构构建多尺度增强特征金字塔(multi-scale enhanced feature pyramid, MSEFP),实现对微小缺陷特征的高效多尺度融合;最后,基于工业现场采集构建医用手套专用缺陷数据集并进行验证。实验结果表明,改进算法PAM-DETR在mAP@50指标上相较于强基线模型YOLO11n和RT-DETR-R18分别提升4.12与4.84个百分点;与RT-DETR-R18相比,PAM-DETR参数量减少14.5%,计算量(FLOPs)降低7.9%,能够较好地满足生产线对医用手套缺陷高精度、轻量化的检测需求。

关键词: 医用手套, 缺陷检测, 小目标检测, RT-DETR, 深度学习

Abstract: To address the issues of low recognition accuracy for tiny targets and strong background interference in medical glove surface defect detection, this paper proposes a lightweight detection model based on the Transformer architecture, named PAM-DETR. First, the PRT-Block module is constructed based on the RT-DETR model, introducing a re-parameterization structure and an attention mechanism. This significantly enhances the representation capability of small target features while eliminating the computational redundancy of standard convolutions. Second, an Adaptive Sparse Encoding Module (ASEM) is designed to replace the original AIFI structure, thereby optimizing the interaction efficiency across different feature scales. Third, by integrating the ASEM with the newly designed CSPOmniKernel structure, a Multi-Scale Enhanced Feature Pyramid (MSEFP) is constructed to achieve efficient multi-scale fusion of tiny defect features. Finally, a dedicated medical glove defect dataset is built based on industrial on-site collection for validation. Experimental results demonstrate that the improved PAM-DETR algorithm achieves an increase of 4.12 and 4.84 percentage points in the mAP@50 metric compared with the strong baseline models YOLO11n and RT-DETR-R18, respectively. Furthermore, compared with RT-DETR-R18, the number of parameters in PAM-DETR is reduced by 14.5%, and the computational cost (FLOPs) is decreased by 7.9%, effectively meeting the production line’s dual requirements for high-precision and lightweight defect detection of medical gloves.

Key words: medical gloves, defect detection, small object detection, RT-DETR, deep learning

中图分类号: TP391.41

王成龙, 宋强, 李文峰, 张仕民. PAM-DETR: 基于改进RT-DETR的医用手套小目标缺陷检测算法[J]. 广西师范大学学报（自然科学版）, 2026, 44(4): 79-95.

Wang Chenglong, Song Qiang, Li Wenfeng, Zhang Shimin. PAM-DETR: a small-object defect detection algorithm for medical gloves based on improved RT-DETR[J]. Journal of Guangxi Normal University(Natural Science Edition), 2026, 44(4): 79-95.

参考文献

[1] GB 24788—2025 医用手套安全技术要求[S].
[2] 万凯. 乳胶制品针孔缺陷检测方法的研究[D]. 南京: 东南大学, 2017.
[3] 李文文, 杨先海, 潘广堂. 一种基于改进Canny算法的塑胶手套残次品检测方法[J]. 塑料科技, 2017, 45(10): 102-105. DOI: 10.15925/j.cnki.issn1005-3360.2017.10.017.
[4] 李嘉欣. 安全套表面缺陷在线检测技术研究[D]. 沈阳: 沈阳工业大学, 2022.
[5] Girshick R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//2014 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos, CA: IEEE Computer Society, 2014: 580-587. DOI: 10.1109/CVPR.2014.81.
[6] Redmon J, Divvala S, Girshick R, et al. You only look once: unified, real-time object detection[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos, CA: IEEE Computer Society, 2016: 779-788. DOI: 10.1109/CVPR.2016.91.
[7] 朱世元, 方世鹏. 基于深度学习的橡胶制品缺陷检测技术创新[J]. 粘接, 2023, 50(7): 26-29. DOI: 10.3969/j.issn.1001-5922.2023.07.007.
[8] 王宇轩. 基于深度学习的手套检测技术研究与应用[D]. 北京: 北方工业大学, 2023.
[9] Lin T Y, Maire M, Belongie S, et al. Microsoft COCO: common objects in context[C]//Computer Vision-ECCV 2014. Cham: Springer, 2014: 740-755. DOI: 10.1007/978-3-319-10602-1_48.
[10] 任书玉, 汪晓丁, 林晖. 目标检测中注意力机制综述[J]. 计算机工程, 2024, 50(12): 16-32.DOI: 10.19678/j.issn.1000-3428.0068553.
[11] Dharma F P, Singgih M L, Prastyo D D. Beyond architecture: hyperparameter optimization for YOLOv8m in multi-class textile defect detection[J]. Results in Engineering, 2025, 28: 108356. DOI: 10.1016/j.rineng.2025.108356.
[12] 刘玉娜, 马双宝. 基于改进YOLOv8n的轻量化织物疵点检测算法[J]. 广西师范大学学报(自然科学版), 2025,43(2): 83-94. DOI: 10.16088/j.issn.1001-6600.2024051302.
[13] Carion N, Massa F, Synnaeve G, et al. End-to-end object detection with transformers[C]//Computer Vision-ECCV 2020: LNCS Volume 12346. Cham: Springer International Publishing AG, 2020: 213-229. DOI: 10.1007/978-3-030-58452-8-21.
[14] Zhu X Z, Su W J, Lu L W, et al. Deformable DETR: deformable transformers for end-to-end object detection[PP/OL]. V4.arXiv(2021-03-18)[2025-11-26]. https://arxiv.org/abs/2010.04159. DOI: 10.48550/arXiv.2010.04159.
[15] Liu S L, Li F, Zhang H, et al. DAB-DETR: dynamic anchor boxes are better queries for DETR[PP/OL]. V4.arXiv(2022-03-30)[2025-11-26]. https://arxiv.org/abs/2201.12329. DOI: 10.48550/arXiv.2201.12329.
[16] Zhao Y A, Lv W Y, Xu S L, et al. DETRs beat YOLOs on real-time object detection[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos, CA: IEEE Computer Society, 2024: 16965-16974. DOI: 10.1109/CVPR52733.2024.01605.
[17] 王东城, 段伯伟, 邢佳文, 等. 基于改进RT-DETR的铜带表面缺陷轻量化检测方法[J]. 中国有色金属学报, 2025, 35(10): 3527-3538. DOI: 10.11817/j.ysxb.1004.0609.2025-45718.
[18] Saeed F, Paul A. ISO-DeTr: a novel detection transformer for industrial small object detection[J]. Machine Learning with Applications, 2026, 23: 100809. DOI: 10.1016/j.mlwa.2025.100809.
[19] Kuhn H W. The Hungarian method for the assignment problem[J]. Naval Research Logistics Quarterly, 1955, 2(1/2): 83-97. DOI: 10.1002/nav.3800020109.
[20] He K M, Zhang X Y, Ren S Q, et al. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos, CA: IEEE Computer Society, 2016: 770-778. DOI: 10.1109/CVPR.2016.90.
[21] 刘耿焕, 曾祥津, 豆嘉真, 等. 基于深度学习的小目标检测技术研究进展[J]. 红外与激光工程, 2024, 53(9): 20240253. DOI: 10.3788/IRLA20240253.
[22] Zhou S H, Chen D S, Pan J S, et al. Adapt or perish: adaptive sparse transformer with attentive feature refinement for image restoration[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos, CA: IEEE Computer Society, 2024: 2952-2963. DOI: 10.1109/CVPR52733.2024.00285.
[23] Chen S, Zhang H Z, Atapour-Abarghouei A, et al. SEM-Net: efficient pixel modelling for image inpainting with spatially enhanced SSM[C]//2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). Los Alamitos, CA: IEEE Computer Society, 2025: 461-471. DOI: 10.1109/WACV61041.2025.00055.
[24] Zhu J C, Chen X L, He K M, et al. Transformers without normalization[PP/OL]. V2.arXiv(2025-06-14)[2025-11-26]. https://arxiv.org/abs/2503.10622. DOI: 10.48550/arXiv.2503.10622.
[25] Yin D S, Hu L Y, Li B, et al. 5%>100%: breaking performance shackles of full fine-tuning on visual recognition tasks[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos, CA: IEEE Computer Society, 2025: 20071-20081. DOI: 10.1109/CVPR52734.2025.01869.
[26] Shazeer N. GLU variants improve transformer[PP/OL]. V1.arXiv(2020-02-12)[2025-11-26]. https://arxiv.org/abs/2002.05202. DOI: 10.48550/arXiv.2002.05202.
[27] Cui Y N, Ren W Q, Knoll A. Omni-Kernel network for image restoration[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38(2): 1426-1434. DOI: 10.1609/aaai.v38i2.27907.
[28] Wang C Y, Liao H M, Wu Y H, et al. CSPNet: a new backbone that can enhance learning capability of CNN[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Los Alamitos, CA: IEEE Computer Society, 2020: 1571-1580. DOI: 10.1109/CVPRW50498.2020.00203.
[29] Sunkara R, Luo T. No more strided convolutions or pooling: a new CNN building block for low-resolution images and small objects[C]//Machine Learning and Knowledge Discovery in Databases: LNCS Volume 13715. Cham: Springer Nature Switzerland AG, 2022: 443-459. DOI: 10.1007/978-3-031-26409-2_27.
[30] Lei M Q, Li S Q, Wu Y H, et al. YOLOv13: real-time object detection with hypergraph-enhanced adaptive visual perception[PP/OL]. V2.arXiv(2025-09-05)[2025-11-26]. https://arxiv.org/abs/2506.17733. DOI: 10.48550/arXiv.2506.17733.
[31] Lou M, Yu Y Z. OverLoCK: an overview-first-look-closely-next ConvNet with context-mixing dynamic kernels[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos, CA: IEEE Computer Society, 2025: 128-138. DOI: 10.1109/CVPR52734.2025.00021.
[32] 胡玉恒, 吴谨. 改进YOLO-DETR的布料表面微小损伤检测方法[J]. 现代电子技术, 2024, 47(13): 160-163. DOI: 10.16652/j.issn.1004-373x.2024.13.028.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed