计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (16): 232-239.DOI: 10.3778/j.issn.1002-8331.2210-0450

• 图形图像处理 • 上一篇    下一篇

融合Transformer的带钢缺陷实时检测算法

张涛源,谢新林,谢刚,张林   

  1. 1.太原科技大学 电子信息工程学院,太原 030024
    2.先进控制与装备智能化山西省重点实验室,太原 030024
    3.平板显示智能制造装备关键技术研发工程研究中心,太原 030024
  • 出版日期:2023-08-15 发布日期:2023-08-15

Real-Time Strip Steel Defect Detection Algorithm Fused with Transformer

ZHANG Taoyuan, XIE Xinlin, XIE Gang, ZHANG Lin   

  1. 1.School of Electronic Information Engineering, Taiyuan University of Science and Technology, Taiyuan 030024, China
    2.Shanxi Key Laboratory of Advanced Control and Equipment Intelligence, Taiyuan 030024, China
    3.Research and Development Engineering Research Center for Key Technology of Flat Panel Display Intelligent Manufacturing, Taiyuan 030024, China
  • Online:2023-08-15 Published:2023-08-15

摘要: 在带钢的生产过程中通常会产生影响产品质量的表面缺陷。针对带钢表面缺陷检测效率低以及小目标缺陷检测精度差的问题,提出一种融合Transformer的带钢缺陷实时检测算法TRSD-YOLO(Transformer real-time strip steel defects detection-YOLO)。设计一种结合Transformer自注意力机制的特征提取模块BottleNeckCSPTR,通过自注意力的增强来提升模块对小目标缺陷信息的获取能力;运用BottleNeckCSPTR模块构建新的主干特征提取网络CSPDarknetTR,并将动态激活函数Meta-ACON与主干网络相融合,进一步强化网络对缺陷特征的表示能力;提出一种轻量级双向加权特征金字塔结构BiFPN-Light作为融合多尺度特征的方式,提高网络对小尺寸缺陷的检测精度。实验结果表明,提出的算法在NEU-DET数据集上mAP达到了82.2%,较原有的YOLOv4算法提高了5.3个百分点;同时检测速度达到31.3?FPS,可匹配工业场景的需求。

关键词: 带钢缺陷检测, YOLOv4, Transformer, 双向特征金字塔(BiFPN)

Abstract: In the production process of strip steel, surface defects often affect the quality of products. Aiming at the problems of poor detection accuracy and low detection efficiency of existing detection methods for small target defects on strip steel surface, a real-time detection algorithm TRSD-YOLO(Transformer real-time strip steel defects detection YOLO) fused with Transformer is proposed. Firstly, a feature extraction module BottleNeckCSPTR combined with Transformer’s self-attention mechanism is designed to enhance the module’s ability to obtain small target defect information through the enhancement of self-attention. Secondly, the BottleNeckCSPTR module is used to build a new backbone feature extraction network CSPDarknetTR, and the dynamic activation function Meta-ACON is integrated with the backbone to further strengthen the network’s ability to represent defect features. Finally, a lightweight bidirectional weighted feature pyramid structure BiFPN-Light is proposed as a way to fuse multi-scale features to improve the network’s detection accuracy for small defects. The experimental results show that the algorithm proposed in this paper achieves a mAP of 82.2% on the NEU-DET dataset, which is 5.3 percentage points higher than the original YOLOv4 algorithm. At the same time, the detection speed reaches 31.3 FPS, which can meet the needs of industrial scenarios.

Key words: strip steel defect detection, YOLOv4, Transformer, bidirectional feature pyramid network(BiFPN)