融合透视校正与轻量注意力机制的轮毂缺陷检测方法

doi:10.11772/j.issn.1001-9081.2025050666

《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (6): 2007-2015.DOI: 10.11772/j.issn.1001-9081.2025050666

• 多媒体计算与计算机仿真 • 上一篇下一篇

融合透视校正与轻量注意力机制的轮毂缺陷检测方法

张纾豪¹^,², 何坤金¹^,²(), 徐佳晨², 沙河山², 陈正鸣¹^,²

^1.海上智能网信技术教育部重点实验室（河海大学），江苏常州 213200
^2.河海大学信息科学与工程学院，江苏常州 213200

收稿日期:2025-06-23 修回日期:2025-07-21 接受日期:2025-07-23 发布日期:2025-08-01 出版日期:2026-06-10
通讯作者: 何坤金
作者简介:张纾豪（2002—），男，云南昆明人，硕士研究生，主要研究方向：深度学习、目标检测
徐佳晨（2001—），男，江苏苏州人，硕士研究生，主要研究方向：图像处理、特征匹配
沙河山（2003—），男，河北邢台人，硕士研究生，主要研究方向：图像处理、软硬件协同设计
陈正鸣（1965—），男，浙江东阳人，教授，博士，CCF高级会员，主要研究方向：计算机辅助设计、计算机图形学。
第一联系人：何坤金（1974—），男，安徽芜湖人，教授，博士，CCF高级会员，主要研究方向：计算机辅助设计、计算机图形学
基金资助:
国家重点研发计划项目(2020YFB1708900);中央高校基本科研业务费专项(B240203012)

Wheel hub defect detection method based on perspective correction and lightweight attention mechanism

Shuhao ZHANG¹^,², Kunjin HE¹^,²(), Jiachen XU², Heshan SHA², Zhengming CHEN¹^,²

^1.Key Laboratory of Maritime Intelligent Cyberspace Technology of Ministry of Education （Hohai University），Changzhou Jiangsu 213200，China
^2.College of Information Science and Engineering，Hohai University，Changzhou Jiangsu 213200，China

Received:2025-06-23 Revised:2025-07-21 Accepted:2025-07-23 Online:2025-08-01 Published:2026-06-10
Contact: Kunjin HE
About author:ZHANG Shuhao， born in 2002， M. S. candidate. His research interests include deep learning， object detection.
XU Jiachen， born in 2001， M. S. candidate. His research interests include image processing， feature matching.
SHA Heshan， born in 2003， M. S. candidate. His research interests include image processing， hardware-software co-design.
CHEN Zhengming， born in 1965， Ph. D.， professor. His research interests include computer-aided design， computer graphics.
First author contact:HE Kunjin， born in 1974， Ph. D.， professor. His research interests include computer-aided design， computer graphics.
Supported by:
National Key Research and Development Program of China(2020YFB1708900);Fundamental Research Funds for the Central Universities(B240203012)

摘要/Abstract

摘要：

工业视觉检测下的轮毂表面缺陷检测任务中，拍摄视角偏差引发的几何畸变，以及缺陷目标尺度小、形态复杂等因素会导致现有检测方法的性能受限。针对上述问题，提出一种融合透视校正与轻量注意力机制的缺陷检测方法。首先，利用椭圆的圆心与轮毂几何中心偏移的关系，构建透视变换四边形，以求解单应性矩阵完成图像视角校正，从而消除畸变对后续特征提取的影响；其次，在YOLOv11模型的基础上，采用轻量化的Ghost卷积替换主干网络与颈部结构中的传统CBS （Convolution-BatchNorm-SiLU（Sigmoid Linear Unit））模块，从而降低模型的参数量与计算量；同时，引入高效通道注意力（ECA）机制，增强网络对微小缺陷区域的感知能力，构建改进模型YOLOv11n-GAConv。实验结果表明，在自建轮毂缺陷数据集上，所提模型预测框与真实框的交并比阈值设为0.5时的平均精度均值（mAP@0.5）达到84.7%，相较于YOLOv11n提升了2.4个百分点，召回率达到79.5%，相较于YOLOv11n 提升了8.6个百分点，而模型的参数量与计算量分别下降12.4%与11.1%。可见，所提方法在提升检测精度的同时降低了模型复杂度。

关键词: YOLOv11, 轮毂缺陷检测, 透视校正, 轻量化结构, 高效通道注意力机制

Abstract:

In wheel hub surface defect detection tasks in industrial visual inspection， geometric distortion caused by shooting angle deviation as well as the small scale and complex morphology of defect objects limit performance of the existing detection methods. To address these challenges， a defect detection method that combines perspective correction and lightweight attention mechanism was proposed. Firstly， the offset relationship between the ellipse center and the geometric center of the wheel hub was utilized to construct a perspective transformation quadrilateral for solving the homography matrix， so as to complete image perspective correction， thereby mitigating the impact of distortion on subsequent feature extraction. Secondly， based on YOLOv11 model， the conventional CBS （Convolution-BatchNorm-SiLU（Sigmoid Linear Unit）） modules in the backbone and neck were replaced by lightweight Ghost convolutions to reduce the number of parameters and the computational cost of the model. Meanwhile， an Efficient Channel Attention （ECA） mechanism was introduced to enhance the network’s perceptual ability to tiny defect regions， leading to the construction of the improved model YOLOv11n-GAConv. Experimental results on a self-built wheel hub defect dataset show that the mean Average Precision of the proposed model when the intersection over union threshold between the predicted bounding boxes and the ground truth bounding boxes is set to 0.5 （mAP@0.5） of 84.7%， with an improvement of 2.4 percentage points compared to that of YOLOv11n， and achieves a recall of 79.5%， with an improvement is 8.6 percentage points compared to that of YOLOv11n. At the same time， the number of parameters and computational cost of the proposed model are reduced by 12.4% and 11.1%， respectively， compared to those of YOLOv11n. It can be seen that the proposed method reduces the model complexity while achieving improved detection precision.

Key words: YOLOv11, wheel hub defect detection, perspective correction, lightweight structure, Efficient Channel Attention (ECA) mechanism

中图分类号:

TP391.41

张纾豪, 何坤金, 徐佳晨, 沙河山, 陈正鸣. 融合透视校正与轻量注意力机制的轮毂缺陷检测方法[J]. 计算机应用, 2026, 46(6): 2007-2015.

Shuhao ZHANG, Kunjin HE, Jiachen XU, Heshan SHA, Zhengming CHEN. Wheel hub defect detection method based on perspective correction and lightweight attention mechanism[J]. Journal of Computer Applications, 2026, 46(6): 2007-2015.

图/表 16

图1 透视变换原理

Fig. 1 Principle of perspective transformation

图2 椭圆中心与轮毂中心

Fig. 2 Ellipse center and wheel hub center

图3 构建透视四边形

Fig.3 Construction of perspective quadrilateral

图4 透视校正前后轮毂的对比

Fig. 4 Comparison of wheel hub before and after perspective correction

图5 YOLOv11n-GAConv模型的结构

Fig. 5 Structure of YOLOv11n-GAConv model

图6 GAConv模块的结构

Fig. 6 Structure of GAConv module

图7 Ghost卷积原理

Fig. 7 Principle of Ghost convolution

图8 ECA模块的结构

Fig. 8 Structure of ECA module

表1 自建轮毂缺陷数据集的标签统计

Tab. 1 Label statistics of self-built wheel hub defect dataset

瑕疵类别	训练集	测试集	平均框面积/px²
杂质	2 230	559	192.9±437.3
条痕	195	48	10 901.1±9 135.2

表2 不同检测模型在自建轮毂缺陷数据集上的性能对比

Tab. 2 Performance comparison of different detection models on self-built wheel hub defect dataset

模型	P/%	R/%	mAP@0.5/%	参数量/10⁶	计算量/GFLOPs	FPS
Faster R-CNN	33.8	65.2	49.2	137.10	370.2	13.4
SSD	67.6	54.2	52.5	26.29	62.8	71.8
RT-DETR-l	66.2	60.0	57.8	31.99	103.4	36.9
YOLOv6n	73.2	56.7	61.1	4.23	11.8	101.1
YOLOv7-tiny	73.5	44.6	52.3	6.23	13.9	71.8
YOLOv9t	70.3	81.7	80.3	1.97	7.6	71.4
YOLOv10n	63.9	76.1	77.5	2.70	8.2	96.2
YOLOv11n	85.6	70.9	82.3	2.58	6.3	100.0
YOLOv12n	87.8	72.9	80.7	2.53	6.0	82.9
YOLOv11n-ShuffleNetV2Unit	80.6	74.2	80.7	2.03	5.7	82.6
YOLOv11n-MobileNetV3Block	81.2	75.8	81.8	2.19	6.0	80.9
YOLOv11n-GAConv	77.8	79.5	84.7	2.26	5.6	86.3

图9 不同模型的Precision-Recall曲线

Fig. 9 Precision-Recall curves of different models

图10 不同检测模型在自建轮毂缺陷数据集上的检测结果可视化对比

Fig. 10 Detection results visual comparison of different models on self-built wheel hub defect dataset

表3 消融实验结果

Tab. 3 Results of ablation experiments

模块结构	替换位置	P/%	R/%	mAP@0.5/%	参数量/10⁶	计算量/GFLOPs
CBS（baseline）		75.3	72.5	78.6	2.58	6.3
Ghost卷积	Backbone+Neck	71.6	74.8	77.9	2.26	5.5
CBS+ECA	Backbone+Neck	75.0	75.6	79.3	2.58	6.4
Ghost卷积+ECA	Backbone	74.0	72.5	77.2	2.35	5.7
Ghost卷积+ECA	Neck	76.2	72.9	78.4	2.49	6.2
Ghost卷积+ECA	Backbone+Neck	77.2	73.6	79.6	2.26	5.6

表4 Ghost卷积线性变换的对比

Tab. 4 Comparison of linear transformations in Ghost convolution

线性变换

函数

mAP@0.5/

表5 注意力机制的对比

Tab. 5 Comparison of attention mechanisms

模块结构	P/%	R/%	mAP@0.5/%	参数量/10⁶	计算量/GFLOPs
Ghost卷积+CBAM	75.8	71.3	78.0	2.37	5.7
Ghost卷积+SE	76.5	75.2	79.4	2.27	5.6
Ghost卷积+SimAM	73.7	73.5	77.6	2.26	5.6

表6 ECA模块的卷积核尺寸对比

Tab. 6 Comparison of convolution kernel sizes in ECA module

k	P/%	R/%	mAP@0.5/%
1	75.3	74.3	77.8
3	77.2	73.6	79.6
5	74.4	73.7	78.4
7	76.6	75.4	79.0
9	73.6	76.9	78.4

参考文献 28

[1]	尚明姝，王克朝，高玉宝. 一种基于SURF-ORB的改进图像配准算法［J］. 计量学报， 2025， 46（3）： 323-328.
	SHANG M S， WANG K C， GAO Y B. An improved image registration algorithm based on SURF-ORB［J］. Acta Metrologica Sinica， 2025， 46（3）： 323-328.
[2]	王才东，刘丰阳，李志航，等. 基于双目视觉特征点匹配的图像拼接方法研究［J］. 激光与光电子学进展， 2021， 58（12）： No.1211002.
	WANG C D， LIU F Y， LI Z H， et al. Research on image mosaic method based on binocular vision［J］. Laser and Optoelectronics Progress， 2021， 58（12）： No.1211002.
[3]	潘宇强，姚垚，张林，等. 基于不规则目标检测的指针式仪表读数识别方法［J］. 仪表技术与传感器， 2024（6）： 100-105.
	PAN Y Q， YAO Y， ZHANG L， et al. Reading recognition method for pointe meters based on irregular object detection［J］. Instrument Technique and Sensor， 2024（6）： 100-105.
[4]	邢浩强，杜志岐，苏波. 变电站指针式仪表检测与识别方法［J］. 仪器仪表学报， 2017， 38（11）： 2813-2821.
	XING H Q， DU Z Q， SU B. Detection and recognition method for pointer-type meter in transformer substation［J］. Chinese Journal of Scientific Instrument， 2017， 38（11）： 2813-2821.
[5]	夏晓华，李玉强，邹易清，等. 圆柱曲面透视投影失真的图像校正［J］. 光学精密工程， 2023， 31（11）： 1691-1699.
	XIA X H， LI Y Q， ZOU Y Q， et al. Image correction for perspective projection distortion of cylindrical surface［J］. Optics and Precision Engineering， 2023， 31（11）： 1691-1699.
[6]	AMERI R， HSU C C， BAND S S. A systematic review of deep learning approaches for surface defect detection in industrial applications［J］. Engineering Applications of Artificial Intelligence， 2024， 130： No.107717.
[7]	WU L， RAN Y， YAN L， et al. A dataset for surface defect detection on complex structured parts based on photometric stereo［J］. Scientific Data， 2025， 12： No.276.
[8]	王素琴，程成，石敏，等. 结合频率和ViT的工业产品表面相似特征缺陷检测方法［J］. 中国图象图形学报， 2024， 29（10）： 3074-3089.
	WANG S Q， CHENG C， SHI M， et al. Defect detection method for industrial product surfaces with similar features by combining frequency and ViT［J］. Journal of Image and Graphics， 2024， 29（10）： 3074-3089.
[9]	郭艳，王智文，赵润星. YOLO-POD：基于多维注意力机制的高精度PCB微小缺陷检测算法［J］. 电子学报， 2024， 52（7）： 2515-2528.
	GUO Y， WANG Z W， ZHAO R X. YOLO-POD： high-precision PCB tiny-defect detection algorithm based on multi-dimensional attention mechanism［J］. Acta Electronica Sinica， 2024， 52（7）： 2515-2528.
[10]	WANG C Y， BOCHKOVSKIY A， MARK LIAO H Y. YOLOv7： trainable bag-of-freebies sets new state-of-the-art for real-time object detectors［C］// Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2023： 7464-7475.
[11]	LIU Y， LIU Y， GUO X， et al. Metal surface defect detection using SLF-YOLO-enhanced YOLOv8 model［J］. Scientific Reports， 2025， 15： No.11105.
[12]	KHANAM R， HUSSAIN M. YOLOv11： an overview of the key architectural enhancements［EB/OL］. ［2025-06-07］..
[13]	HAN K， WANG Y， TIAN Q， et al. GhostNet： more features from cheap operations［C］// Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2020： 1577-1586.
[14]	WANG Q， WU B， ZHU P， et al. ECA-Net： efficient channel attention for deep convolutional neural networks ［C］// Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2020： 11531-11539.
[15]	WANG C， DING Y， CUI K， et al. A perspective distortion correction method for planar imaging based on homography mapping ［J］. Sensors， 2025， 25（6）： No.1891.
[16]	HE Y， SONG K， MENG Q， et al. An end-to-end steel surface defect detection approach via fusing multiple hierarchical features［J］. IEEE Transactions on Instrumentation and Measurement， 2020， 69（4）：1493-1504.
[17]	REN S， HE K， GIRSHICK R， et al. Faster R-CNN： towards real-time object detection with region proposal networks［C］// Proceedings of the 28th International Conference on Neural Information Processing Systems — Volume 1. Cambridge： MIT Press， 2015： 91-99.
[18]	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.
[19]	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.
[20]	LI C， LI L， JIANG H， et al. YOLOv6： a single-stage object detection framework for industrial applications ［EB/OL］. ［2025-06-07］. .
[21]	WANG C Y， YEH I H， MARK LIAO H Y. YOLOv9： learning what you want to learn using programmable gradient information［C］// Proceedings of the 2024 European Conference on Computer Vision， LNCS 15089. Cham： Springer， 2025： 1-21.
[22]	WANG A， CHEN H， LIU L， et al. YOLOv10： real-time end-to-end object detection ［C］// Proceedings of the 38th Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2024： 107984-108011.
[23]	TIAN Y， YE Q， DOERMANN D. YOLOv12： attention-centric real-time object detectors［EB/OL］. ［2025-06-07］..
[24]	MA N， ZHANG X， ZHENG H T， et al. ShuffleNet V2： practical guidelines for efficient CNN architecture design ［C］// Proceedings of the 2018 European Conference on Computer Vision， LNCS 11218. Cham： Springer， 2018： 122-138.
[25]	HOWARD A， SANDLER M， CHEN B， et al. Searching for MobileNetV3［C］// Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision. Piscataway： IEEE， 2019： 1314-1324.
[26]	WOO S， PARK J， LEE J Y， et al. CBAM： convolutional block attention module［C］// Proceedings of the 2018 European Conference on Computer Vision， LNCS 11211. Cham： Springer， 2018： 3-19.
[27]	HU J， SHEN L， SUN G. Squeeze-and-excitation networks［C］// Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2018： 7132-7141.
[28]	YANG L， ZHANG R Y， LI L， et al. SimAM： a simple， parameter-free attention module for convolutional neural networks［C］// Proceedings of the 38th International Conference on Machine Learning. New York： JMLR.org， 2021： 11863-11874.

融合透视校正与轻量注意力机制的轮毂缺陷检测方法

Wheel hub defect detection method based on perspective correction and lightweight attention mechanism

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 16

参考文献 28

相关文章 2

编辑推荐

Metrics

[1]	尹秋燕, 丁婧, 聂志刚. 无人机航拍视角下的人体姿态估计算法YOLO-AirPose[J]. 《计算机应用》唯一官方网站, 2026, 46(6): 1989-1997.
[2]	杜艺, 续明进, 孔佳仪, 王力瑶, 赵晨. 基于YOLOv11的低秩自适应参数高效微调算法[J]. 《计算机应用》唯一官方网站, 2026, 46(6): 1738-1745.