钢板表面缺陷智能检测技术演进、关键算法与工业化落地研究

Research on Evolution, Key Algorithms and Industrial Implementation of Intelligent Steel Plate Surface Defect Detection Technology

摘要: 钢板是制造业重要基础原料，其表面裂纹、夹杂等缺陷会造成构件应力集中，显著降低产品服役安全性能。传统人工质检效率低下，漏检、误检问题突出，无法适配高速轧制产线的生产节拍，基于机器视觉与深度学习的钢板表面缺陷智能检测，已成为钢铁智能制造的重要研究方向。本文采用综述–原创算法–工业实测的复合型研究框架，针对小目标缺陷漏检、边缘设备算力受限、单模态成像识别能力不足三大行业痛点，梳理三代缺陷检测技术发展路线；基于NEU-DET公开数据集开展多算法量化对比实验，对比不同算法的mAP、参数量、FPS性能指标；剖析轻量化单阶段网络、CNN-Transformer融合、2D/3D复合成像、GAN样本扩充四项关键技术；结合热轧卷材、厚板两条钢厂实际产线项目，给出缺陷检出率、人力成本节约等工业实测数据；探讨样本不均衡、模型黑盒特性、中小企业设备投入成本等工程落地难题，并展望行业未来发展方向。原创贡献：1) 构建GhostNet + ShuffleNetv2双骨干YOLOv8-SP网络，在NEU-DET数据集上mAP达到82.3%，相比YOLOv8-n提升5.2%；2) 搭建基于MobileNetv3的YOLOv8-SPL轻量化模型，参数量仅1.89  M，推理速度326  FPS，可满足边缘工控设备部署要求；3) 提出2D/3D联合成像完整落地方案与标准化调试流程，为钢板表面缺陷检测装备工程应用提供实测参考。

Abstract: As an important basic raw material for manufacturing, surface defects such as cracks and inclusions on steel plates will cause stress concentration of components and significantly reduce the service safety performance of products. Traditional manual inspection suffers from low efficiency and frequent missed-false detection, which cannot match the production rhythm of high-speed rolling lines. Intelligent steel-plate defect detection based on machine vision and deep learning has become a vital research direction for iron-and-steel intelligent manufacturing. Adopting a composite framework of literature review, original algorithm design and industrial measurement, this paper addresses three major industrial bottlenecks: missed detection of small-target defects, limited computing power of edge devices, and insufficient recognition capability of single-modal imaging. The development routes of three generations of defect‑detection technologies are sorted out. Quantitative comparison experiments of multiple algorithms are conducted on the NEU-DET public dataset to compare their mAP, parameter quantity and FPS indicators. Four key technologies are analyzed, including lightweight single‑stage network, CNN-Transformer fusion, 2D/3D composite imaging and GAN sample augmentation. Combined with two practical steel-plant projects of hot-rolled coils and heavy plates, industrial measured data such as defect detection rate and labor-cost saving are provided. Practical engineering challenges including sample imbalance, model black-box property and equipment investment cost for small‑and‑medium enterprises are discussed, and future development trends of this field are prospected. Original Contributions: 1) The GhostNet+ShuffleNetv2 dual-backbone YOLOv8-SP network is constructed, achieving 82.3% mAP on the NEU-DET dataset, 5.2% higher than YOLOv8-n. 2) The MobileNetv3-based lightweight YOLOv8-SPL model is established, with only 1.89 M parameters and an inference speed of 326 FPS, suitable for deployment on edge industrial control devices. 3) A complete deployment scheme and standardized debugging procedure for 2D/3D joint imaging are proposed, providing measured references for the engineering application of steel-plate defect-detection equipment.