Defect detection algorithm for train bearing rollers based on FHC-DETR

doi:10.11772/j.issn.1001-9081.2025050592

Abstract

Abstract:

To address the issues such as low precision in small object recognition and large-scale variations in the detection of train bearing roller defects， a detection algorithm based on FHC-DETR（Fourier-fused High-low frequency interactive and Context-aware DEtection TRansformer） was proposed. Firstly， aiming at the problems of complex model computation and small object features being disturbed by noise， a frequency-domain feature extraction module C2f-FG （C2f with Fourier-Gated bottleneck） was designed for feature extraction. It synchronously acquired spatial-domain local features and frequency-domain global features via Fourier transform， and their fusion enhanced the accuracy of small-object detection while reducing computational complexity. Secondly， to tackle feature confusion caused by variations in defect scale， a high-low-frequency feature interaction module， HiLo （High-Low frequency）， was introduced. The high-frequency branch focused on local defect textures， while the low-frequency branch captured overall semantics via global attention， thereby improving multi-scale adaptability. Subsequently， to resolve the issue of small object feature attenuation in feature fusion， a Context-aware Multi-scale Bidirectional Feature Pyramid Network （CM-BiFPN） was constructed. By dynamically perceiving context and strengthening cross-layer interaction， it reduced feature transmission loss and improved fusion efficiency. Finally， the EMASlideVarifocalLoss adaptive loss function was adopted to dynamically adjust classification thresholds and optimize weights of hard examples， further enhancing localization and category discrimination capabilities. Experimental results show that FHC-DETR achieves a mean Average Precision （mAP） of 91.5%， which is 2.3 percentage points higher than that of the original RT-DETR （Real-Time DETR）. Additionally， its parameter count is reduced by 28.1%， its computational load is reduced by 23.0%， and its memory usage is reduced by 23.7%， demonstrating a balance between precision and efficiency and confirming its practicality in industrial scenarios.

Key words: defect detection, small object, RT-DETR (Real-Time DEtection TRansformer), Feature Pyramid Network (FPN), loss function

摘要：

针对列车轴承滚子缺陷检测中存在的小目标识别精度低和尺度变化大等问题，提出基于FHC-DETR（Fourier-fused High-low frequency interactive and Context-aware DEtection TRansformer）的检测算法。首先，针对模型计算复杂及小目标特征受噪声干扰的问题，设计频域特征提取模块C2f-FG（C2f with Fourier-Gated bottleneck）提取特征，通过傅里叶变换同步获取空间域局部特征与频域全局特征，二者融合可提高小目标检测精度并降低计算复杂度；其次，针对缺陷尺度变化导致的特征混淆，引入高低频特征交互模块HiLo（High-Low frequency），高频分支聚焦局部缺陷纹理，低频分支通过全局注意力捕捉整体语义，从而提升多尺度适应能力；再次，针对特征融合中存在的小目标特征衰减问题，构建上下文感知型多尺度双向特征金字塔网络（CM-BiFPN），通过动态感知上下文并强化跨层交互，减小特征传递损失，提升融合效率；最后，采用EMASlideVarifocalLoss自适应损失函数，动态调整分类阈值并优化难例权重，进一步提升定位与类别区分能力。实验结果显示，所提检测算法的平均精度均值（mAP）达91.5%，比实时检测变换器（RT-DETR）高2.3个百分点，且参数量减少28.1%，计算量降低23.0%，内存占用下降23.7%，实现了精度与效率的平衡，验证了它在工业场景的实用性。

关键词: 缺陷检测, 小目标, 实时检测变换器, 特征金字塔网络, 损失函数

CLC Number:

TP391.41

Yuanhao HE, Jun ZHAO. Defect detection algorithm for train bearing rollers based on FHC-DETR[J]. Journal of Computer Applications, 2026, 46(5): 1624-1633.

贺元昊, 赵军. 基于FHC-DETR的列车轴承滚子缺陷检测算法[J]. 《计算机应用》唯一官方网站, 2026, 46(5): 1624-1633.

Figures/Tables 21

References 29

[1]	ZHOU P， ZHANG C， ZHOU G， et al. Whole surface defect detection method for bearing rings based on machine vision［J］. Measurement Science and Technology， 2023， 34（1）： No.015017.
[2]	魏利胜，丁坤，段志达，等.融合单应性约束SIFT特征匹配的轴承滚子检测［J］.电子测量与仪器学报，2019，33（9）：107-113.
	WEI L S， DING K， DUAN Z D， et al. Bearing roller detection based on SIFT feature matching with homography constraint［J］. Journal of Electronic Measurement and Instrumentation， 2019， 33（9）： 107-113.
[3]	陈硕，林志敏，吴岳彬，等.轴承套圈端面缺陷在线视觉检测的研究与实现［J］.轴承，2022（2）：48-54.
	CHEN S， LIN Z M， WU Y B， et al. Research and implementation of on-line visual inspection of defects on end face of bearing rings［J］. Bearing， 2022（2）： 48-54.
[4]	石炜，刘松，杨锦春，等.列车滚子轴承表面缺陷图像的边缘提取［J］.机械设计与制造，2024（11）：179-183.
	SHI W， LIU S， YANG J C， et al. Edge extraction of surface defect image of train roller bearing［J］. Machinery Design and Manufacture， 2024（11）： 179-183.
[5]	张胜华，龙嘉健，陆继东.基于Mask R-CNN的锅炉火焰图像核心高温区域提取［J］.发电设备，2024， 38（4）：205-210.
	ZHANG S H， LONG J J， LU J D. Extraction of core high temperature region of boiler flame image based on Mask R-CNN［J］. Power Equipment， 2024， 38（4）： 205-210.
[6]	周诗捷，王玉槐，沈思橙，等.基于改进型Faster R-CNN的仓储环境物体识别技术研究［J］.计算技术与自化，2024，43（2）：187-191.
	ZHOU S J， WANG Y H， SHEN S C， et al. Research on object recognition in warehouse environment based on improved Faster R-CNN［J］. Computing Technology and Automation， 2024， 43（2）： 187-191.
[7]	MOHAMMED S， NIMALI T M. A new paradigm for waste classification based on YOLOv5［J］. Instrumentation， 2021， 8（4）： 9-17.
[8]	LIU W， ANGUELOV D， ERHAN D， et al. SSD： single shot multiBox detector［C］// Proceedings of the 2016 European Conference on Computer Vision， LNCS 9905. Cham： Springer， 2016： 21-37.
[9]	李晓宇，张涛，刘宏伟，等.基于改进Faster R-CNN的轴承端面缺陷检测方法［J］.计算机仿真，2025，42（1）：491-496.
	LI X Y， ZHANG T， LIU H W， et al. Bearing end face defect detection method based on improved Faster R-CNN［J］. Computer Simulation， 2025， 42（1）： 491-496.
[10]	WANG M， YANG W， WANG L， et al. FE-YOLOv5： feature enhancement network based on YOLOv5 for small object detection［J］. Journal of Visual Communication and Image Representation， 2023， 90： No.103752.
[11]	安冬，胡荣华，王丽艳，等.基于Transformer的陶瓷轴承表面缺陷检测方法［J］.组合机床与自动化加工技术，2024（2）：160-163， 168.
	AN D， HU R H， WANG L Y， et al. Ceramic bearing surface defect detection method based on Transformer［J］. Modular Machine Tool and Automatic Manufacturing Technique， 2024（2）： 160-163， 168.
[12]	TONG K， WU Y. Small object detection using deep feature learning and feature fusion network［J］. Engineering Applications of Artificial Intelligence， 2024， 132： No.107931.
[13]	袁天乐，袁巨龙，朱勇建，等.基于改进YOLOv5的推力球轴承表面缺陷检测算法［J］.浙江大学学报（工学版），2022，56（12）： 2349-2357.
	YUAN T L， YUAN J L， ZHU Y J， et al. Surface defect detection algorithm of thrust ball bearing based on improved YOLOv5［J］. Journal of Zhejiang University （Engineering Science）， 2022， 56（12）： 2349-2357.
[14]	周得辉，赵军，程进峰.基于RT-DETR的轴承表面微小缺陷检测算法［J］.计算机应用，2025，45（6）：1987-1997.
	ZHOU D H， ZHAO J， CHENG J F. Tiny defect detection algorithm for bearing surface based on RT-DETR［J］. Journal of Computer Applications， 2025， 45（6）： 1987-1997.
[15]	李亚东，马行，穆春阳，等.改进YOLOX网络的轴承缺陷小目标检测方法［J］.计算机工程与应用，2023，59（1）：100-107.
	LI Y D， MA X， MU C Y， et al. Improved small target detection method of bearing defects in YOLOX network［J］. Computer Engineering and Applications， 2023， 59（1）： 100-107.
[16]	WANG G， CHEN Y， AN P， et al. UAV-YOLOv8： a small-object-detection model based on improved YOLOv8 for UAV aerial photography scenarios［J］. Sensors， 2023， 23（16）： No.7190.
[17]	ZHAO Y， LV W， XU S， et al. DETRs beat YOLOs on real-time object detection［C］// Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2024： 16965-16974.
[18]	GAO N， JIANG X， ZHANG X， et al. Efficient frequency-domain image deraining with contrastive regularization［C］// Proceedings of the 2024 European Conference on Computer Vision， LNCS 15099. Cham： Springer， 2025： 240-257.
[19]	SHI D. TransNeXt： robust foveal visual perception for Vision Transformers［C］// Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2024： 17773-17783.
[20]	VASWANI A， SHAZEER N， PARMAR N， et al. Attention is all you need［C］// Proceedings of the 31st International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2017：6000-6010.
[21]	PAN Z， CAI J， ZHUANG B. Fast Vision Transformers with HiLo attention［C］// Proceedings of the 36th International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2022： 14541-14554.
[22]	LIU S， QI L， QIN H， et al. Path aggregation network for instance segmentation［C］// Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2018： 8759-8768.
[23]	DOHERTY J， GARDINER B， KERR E， et al. BiFPN-YOLO： one-stage object detection integrating bi-directional feature pyramid networks［J］. Pattern Recognition， 2025， 160： No.111209.
[24]	CAI X， LAI Q， WANG Y， et al. Poly kernel inception network for remote sensing detection［C］// Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2024： 27706-27716.
[25]	DING X， ZHANG X， HAN J， et al. Diverse branch block： building a convolution as an inception-like unit［C］// Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2021： 10881-10890.
[26]	YU Z， HUANG H， CHEN W， et al. YOLO-FaceV2：a scale and occlusion aware face detector［J］. Pattern Recognition， 2024， 155： No.110714.
[27]	LIU Z， WANG Y， VAIDYA S， et al. KAN： Kolmogorov-Arnold networks［EB/OL］. ［2025-04-11］..
[28]	ZHANG S， WANG X， WANG J， et al. Dense distinct query for end-to-end object detection［C］// Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2023： 7329-7338.
[29]	SONG K， YAN Y. A noise robust method based on completed local binary patterns for hot-rolled steel strip surface defects［J］. Applied Surface Science， 2013， 285（Pt B）： 858-864.

名称	参数	名称	参数
操作系统	Ubuntu18.04	优化器	AdamW
CPU	Intel Xeon Gold 6226R	迭代次数	200
GPU	NVIDIA Quadro RXT 5000	批处理大小	8

名称	参数	名称	参数
操作系统	Ubuntu18.04	优化器	AdamW
CPU	Intel Xeon Gold 6226R	迭代次数	200
GPU	NVIDIA Quadro RXT 5000	批处理大小	8

实验	C2f-FG	AIFI-HiLo	CM-BiFPN	EMASlideVarifocalLoss	精确率/%	召回率/%	mAP/%	GFLOPs	参数量/10⁶	内存占用量/MB
Ⅰ					89.1	86.4	89.2	57.0	19.9	38.9
Ⅱ	√				90.9	87.0	90.1	51.8	16.9	33.0
Ⅲ		√			90.5	87.4	89.8	57.1	19.8	38.6
Ⅳ			√		91.5	88.1	90.6	48.2	17.2	35.2
Ⅴ				√	91.1	87.9	90.0	57.0	19.9	38.9
Ⅵ	√	√			91.8	87.5	90.4	52.0	16.9	32.9
Ⅶ	√	√	√		91.9	86.9	91.1	43.9	14.3	29.6
Ⅷ	√	√	√	√	91.4	87.9	91.5	43.9	14.3	29.7

实验	C2f-FG	AIFI-HiLo	CM-BiFPN	EMASlideVarifocalLoss	精确率/%	召回率/%	mAP/%	GFLOPs	参数量/10⁶	内存占用量/MB
Ⅰ					89.1	86.4	89.2	57.0	19.9	38.9
Ⅱ	√				90.9	87.0	90.1	51.8	16.9	33.0
Ⅲ		√			90.5	87.4	89.8	57.1	19.8	38.6
Ⅳ			√		91.5	88.1	90.6	48.2	17.2	35.2
Ⅴ				√	91.1	87.9	90.0	57.0	19.9	38.9
Ⅵ	√	√			91.8	87.5	90.4	52.0	16.9	32.9
Ⅶ	√	√	√		91.9	86.9	91.1	43.9	14.3	29.6
Ⅷ	√	√	√	√	91.4	87.9	91.5	43.9	14.3	29.7

应用次数	精确率/%	召回率/%	mAP/%	GFLOPs	参数量/10⁶
0	89.1	86.4	89.2	57.0	19.9
1	89.6	87.2	89.6	52.4	15.5
2	89.5	86.4	89.7	52.1	15.6
3	89.7	86.5	89.9	51.9	15.9
4	90.9	87.0	90.1	51.8	16.9