| 引用本文: | 周孟然a,蔡 睿b,范桃春b,王 宁b.基于尺度多样化改进的钢材表面缺陷检测算法(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2026，43（3）：70-80 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| :目的 钢材表面缺陷一直是工业生产和安全的难题之一,为解决钢材表面缺陷检测精度低,容易出现误检和
漏检的问题,提出了一种基于多尺度特征的钢材表面缺陷检测算法,命名为 GPDN-Yolov5。 方法 首先,针对部分
缺陷容易与背景相混淆的问题,通过在骨干网络中加入改进后的坐标注意力机制(GJ-CA),加强对缺陷特征的关
注。 其次,针对钢材表面缺陷具有尺寸大小不一的特点,引入多尺度卷积块 PKI Block,加强了网络的多尺度特征
提取能力。 然后,采用 Dynamic head 检测头作为改进算法的检测头,增强了检测头的表达能力和对多尺度特征的
检测能力。 随后,将 NWD(Normalized Wasserstein Distance)损失函数与 WIOU 损失函数加权融合来替代 CIOU 损
失函数,解决了原模型专注于简单样本和对低像素缺陷位置偏差敏感的问题。 结果 GPDN-Yolov5 算法在 NEUDET 数据集中的平均精度均值达到 80. 5%,比 Yolov5 算法提升了 6. 9%,检测结果优于其他经典检测模型。 同时,
该算法在 GC10-DET 数据集上的平均精度比 Yolov5 提升了 4. 7%,这证明了算法在不同数据集上检测效果均有提
升。 结论 GPDN-Yolov5 算法在精度上明显提高,对钢材表面缺陷检测具有一定价值。 |
| 关键词: Yolov5 缺陷检测 注意力机制 多尺度特征 Dynamic head NWD-WIOU |
| DOI： |
| 分类号: |
| 基金项目: |
|
| An Improved Steel Surface Defect Detection Algorithm Based on Multi-Scale Features Algorithm Based on Dual-Feature Fusion |
|
ZHOU Mengrana，CAI Ruib，FAN Taochunb，WANG Ningb
|
|
a. School of Electrical and Information Engineering b. School of Mechanics Optoelectronics and Physics Anhui
University of Science and Technology Huainan 232001 Anhui China of Science and Technology Huainan 232000 Anhui China
|
| Abstract: |
| Objective Surface defects in steel remain a critical challenge for industrial production and safety. To address
the problems of low accuracy false detection and missed detection in steel surface defect detection a steel surface defect
detection algorithm based on multi-scale features named GPDN-Yolov5 is proposed. Methods First to tackle the
problem of certain defects blending with the background an improved coordinate attention mechanism GJ-CA was
incorporated into the backbone network to enhance focus on defect features. Second considering the wide variation in
defect sizes a multi-scale convolutional block PKI Block was introduced to strengthen the network?? s ability to extract
multi-scale features. Then a Dynamic head was applied as the detection head of the improved algorithm which
strengthened the expression ability of the detection head and the detection ability of multi-scale features. Finally the
CIOU loss function was replaced by a weighted combination of the normalized Wasserstein distance NWD loss function and the WIOU loss function thereby addressing the original model?? s bias toward easy samples and its high sensitivity to
positional deviation in low-pixel defects. Results The GPDN-Yolov5 algorithm achieves a mean average precision mAP
of 80. 5% on the NEU-DET dataset representing a 6. 9% improvement over the baseline Yolov5 algorithm and
outperforms other classical detection models. Moreover on the GC10-DET dataset the algorithm improves the mean
average precision by 4. 7% over Yolov5 demonstrating its consistent performance gain across different datasets.
Conclusion The GPDN-Yolov5 algorithm provides a significant improvement in detection accuracy and holds certain value
for steel defect detection. |
| Key words: Yolov5 defect detection attention mechanism multi-scale feature Dynamic head NWD-WIOU |
