基于YOLO系列模型的瓦楞纸箱损伤检测方法研究
Research on YOLO-Series Methods for Corrugated Box Damage Detection

为应对物流现场中瓦楞纸箱因划痕、破洞与湿渍等造成的包装失效问题，本文构建包含3类典型损伤、共3800张图像的多源数据集(实验室模拟、网络采集、驿站实拍)，并在统一训练/评测流程下，对YOLOv3-tiny、YOLOv5n、YOLOv8n、YOLO10n、YOLO11n与YOLO12n六个版本进行系统对比。采用P、R、F1、mAP₅₀与GFLOPs等指标综合评估，并通过混淆矩阵分析不同模型在细粒度损伤上的识别差异。结果显示：YOLOv8n在精度与召回之间取得最优平衡(P = 0.840, R = 0.790, mAP₅₀ = 0.868, F1 = 0.814)，适合作为精度优先的工业检测基线；YOLO12n在保持较高精度的同时具备最低计算复杂度(GFLOPs = 6.3, mAP₅₀ = 0.858)，更适合端侧实时部署。对比分析还表明，弱纹理/反光干扰场景下“wet”与背景的混淆最为突出，而YOLO10n在“hole”类的召回存在明显下降。本文的贡献在于：(i) 提出覆盖多场景的纸箱损伤数据集与统一评测协议；(ii) 给出跨代YOLO的可复现对比基准；(iii) 提供面向工程落地的模型选型建议。未来工作将聚焦于损伤分级标准化、小样本增强与轻量注意力机制，以进一步提升复杂环境下的鲁棒性与可部署性。
To address packaging failures of corrugated boxes caused by scratches, holes, and wet stains in logistics, we curate a 3-class, 3,800-image dataset from laboratory simulation, web crawling, and field acquisition at parcel stations. Under a unified training and evaluation protocol, we benchmark six YOLO variants—YOLOv3-tiny, YOLOv5n, YOLOv8n, YOLO10n, YOLO11n, and YOLO12n. Comprehensive evaluation employing metrics including P, R, F1, mAP50 and GFLOPs, alongside confusion matrix analysis to assess the identification discrepancies among different models regarding fine-grained damage. Results show that YOLOv8n achieves the best accuracy-recall balance (P = 0.840, R = 0.790, mAP₅₀ = 0.868, F1 = 0.814), making it a strong baseline for accuracy-oriented inspection. YOLO12n attains the lowest computational cost (GFLOPs = 6.3) while maintaining competitive accuracy (mAP₅₀ = 0.858), thus favoring real-time edge deployment. We also find noticeable confusion between “wet” and background under weak-texture/reflective conditions, and degraded recall for “hole” with YOLO10n. Our contributions include: (i) a multi-scenario corrugated-box defect dataset with a unified evaluation protocol; (ii) a reproducible cross-generation YOLO benchmark; and (iii) actionable guidance for engineering deployment. Future work will explore standardized severity grading, small-sample augmentation, and lightweight attention to further improve robustness and deployability in the wild.