中文核心期刊
CSCD来源期刊
中国科技核心期刊
RCCSE中国核心学术期刊
重庆交通大学学报（自然科学版）

重庆交通大学学报（自然科学版） ›› 2026, Vol. 45 ›› Issue (9): 110-117.DOI: 10.3969/j.issn.1674-0696.2026.09.13

• 现代交通装备 • 上一篇    

基于机器视觉的钢材表面缺陷检测优化研究

董华军1，杨明翰1，李籽骁1，董华师2，李金金3   

  1. (1. 大连交通大学 机械工程学院，辽宁 大连 116028； 2. 浙江正泰仪器仪表有限责任公司，浙江 温州 325603； 3. 大连交通大学 电气工程学院，辽宁 大连 116028)
  • 收稿日期:2026-03-02 修回日期:2026-08-19 发布日期:2026-09-23
  • 作者简介:董华军(1978—)，男，湖北黄冈人，教授，博士，主要从事真空开关理论研究、图像处理技术及识别方面的研究。E-mail：huajundong4025@163.com 通信作者：李金金(1993—)，男，辽宁盘山人，讲师，博士，主要从事图像处理、机器视觉、高压电器等方面的研究；E-mail：a379415016@163.com
  • 基金资助:
    辽宁省自然科学基金项目（2024-BS-200）；辽宁省属本科高校基本科研业务费专项资金项目（LJ212410150060）

Optimization of steel surface defect detection based on machine vision

Dong Huajun1, Yang Minghan1, Li Zixiao1, Dong Huashi2, Li Jinjin3   

  1. (1. School of Mechanical Engineering, Dalian Jiaotong University, Dalian 116028, Liaoning, China; 2. Zhejiang Chint Instrument & Meter Co., Ltd., Wenzhou 325603, Zhejiang, China; 3. School of Electrical Engineering, Dalian Jiaotong University, Dalian 116028, Liaoning, China)
  • Received:2026-03-02 Revised:2026-08-19 Published:2026-09-23

PDF

摘要： 针对现有钢材表面缺陷检测方法存在精度不足、参数冗余、推理速度慢等问题，提出了一种LSD-YOLOv11高精度检测算法。笔者算法采用FRLDS下采样方法，降低了特征图分辨率，保留了微小缺陷的关键特征，并设计了C3K2-Di-SpAM模块增强多尺度特征提取能力，引入了依赖图的DepGraph结构化剪枝算法，剔除冗余参数。实验结果表明：LSD-YOLOv11检测算法有效平衡了特征分辨率降低与关键信息保留的矛盾，减少了计算量，提升了特征提取能力；相较于基准模型YOLOv11，笔者算法处理结果的mAP50值提高了8.5%，计算量下降了18.9%，FPS提高了21.7%。

关键词: 机械工程；钢材缺陷检测；机器视觉；YOLOv11；算法优化

Abstract: To address the problems such as insufficient accuracy, parameter redundancy and slow inference speed in existing steel surface defect detection methods, a high-precision detection algorithm named LSD-YOLOv11 was proposed. In the proposed algorithm, the FRLDS downsampling method was adopted to reduce the resolution of the feature map and preserve the key features of small defects. The C3K2-Di-SpAM module was designed to enhance the multi-scale feature extraction capability, and the DepGraph structured pruning algorithm based on dependency graphs was introduced to remove redundant parameters. Experimental results show that LSD-YOLOv11 effectively balances the contradiction between feature resolution reduction and key information preservation, reduces the computational cost and improves the feature extraction capability. Compared to the baseline model YOLOv11n, the proposed algorithm improves mAP50 by 8.5%, reduces computational cost by 18.9% and increases FPS by 21.7%.

Key words: mechanical engineering; steel defect detection; machine vision; YOLOv11; algorithm optimization

中图分类号: