基于YOLOv8s的钢材表面缺陷检测方法研究

Steel Surface Defect Detection Based on YOLOv8s

摘要: 钢材表面缺陷的准确检测是保障钢铁产品质量的关键环节。针对热轧带钢表面缺陷种类多样、部分类别与背景纹理相似导致漏检率高等问题，本文针对NEU-DET数据集对YOLOv8s进行系统的优化与评估，构建强有力的基准检测模型。以东北大学热轧带钢表面缺陷数据集(NEU-DET)为实验基础，该数据集涵盖裂纹、夹杂物、斑块、麻面、轧入氧化皮和划痕6类缺陷，共1800张图像。通过高分辨率输入(800 × 800像素)、AdamW优化器、Mosaic/MixUp/CopyPaste多策略数据增强及余弦学习率调度等组合优化策略，相比YOLOv8s默认配置(640 × 640，默认数据增强)显著提升了模型性能。实验结果表明，优化后模型在验证集上mAP@0.5达到0.739 (基线0.701)，较基线提升了3.8%。其中斑块类AP最高为0.923，划痕类为0.868，最优F1分数为0.69 (置信度阈值0.292)。混淆矩阵及可视化分析揭示了各类缺陷的检测特性，为后续改进指明了方向。

Abstract: Accurate detection of steel surface defects is essential for quality assurance in steel manufacturing. Aiming at the problems of diverse defect types and a high miss-detection rate caused by similarity between certain defect categories and background textures in hot-rolled strip steel, this paper systematically optimizes and evaluates YOLOv8s on the NEU-DET dataset to construct a strong baseline detection model. Experiments are conducted on the NEU-DET dataset released by Northeastern University, which covers six types of defects—crazing, inclusion, patches, pitted surface, rolled-in scale, and scratches—totaling 1800 images. Through a combination of high-resolution input (800 × 800 pixels), AdamW optimizer, multi-strategy data augmentation (Mosaic/MixUp/CopyPaste), and cosine learning rate scheduling, the optimized model achieves mAP@0.5 = 0.739 on the validation set, representing a +3.8% improvement over the YOLOv8s default configuration baseline (640 × 640, mAP@0.5 = 0.701). Among individual categories, patches achieve the highest AP of 0.923 and scratches reach 0.868, while the optimal F1 score is 0.69 at a confidence threshold of 0.292. Confusion matrix and visualization analyses reveal the detection characteristics of each defect type and indicate directions for future improvement.