计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (15): 318-328.DOI: 10.3778/j.issn.1002-8331.2303-0100

• 工程与应用 • 上一篇    下一篇

面向复杂环境中带钢表面缺陷检测的轻量级DCN-YOLO

卢俊哲,张铖怡,刘世鹏,宁德军   

  1. 1.中国科学院 上海高等研究院,上海 201210
    2.中国科学院大学,北京 100049
  • 出版日期:2023-08-01 发布日期:2023-08-01

Lightweight DCN-YOLO for Strip Surface Defect Detection in Complex Environments

LU Junzhe, ZHANG Chengyi, LIU Shipeng, NING Dejun   

  1. 1.Shanghai Advanced Research Institute, Chinese Academy of Sciences, Shanghai 201210, China
    2.University of Chinese Academy of Sciences, Beijing 100049, China
  • Online:2023-08-01 Published:2023-08-01

摘要: 基于深度学习的智能检测技术逐渐在复杂钢铁生产环境带钢表面缺陷检测过程中使用。为了应对在资源受限的边缘设备中部署高精度模型的挑战,提出一个面向复杂环境中带钢表面缺陷检测的轻量级DCN-YOLO模型,该模型将可形变卷积网络DCN与原始YOLOv5结合,以提高模型对不同尺寸和形状缺陷的灵敏度。为降低计算复杂度,在YOLO模型中引入了深度可分离卷积DSConv和高效通道注意力机制ECA两个轻量级模块,使模型更好地理解输入数据中各个通道之间的关系,在提高模型的检测精度和泛化能力的同时,大幅降低模型的计算量。进一步通过消融实验及横向对比实验,验证了每个创新模块的有效性。通过经典的开源带钢数据集NEU-DET和实际工业带钢数据集分别验证了轻量级DCN-YOLO模型在表面缺陷检测精度和计算复杂度方面的优势。

关键词: 带钢表面缺陷检测, 可形变卷积网络, 深度可分离卷积, ECA通道注意力, 轻量级YOLOv5, 图像预处理

Abstract: At present, deep learning-based intelligent defect detection technologies has gradually penetrated the traditional steel manufacturing industry. To address the challenge of deploying high-precision models in resource constrained edge devices, a lightweight DCN-YOLO model for strip surface defect detection in complex environment is proposed. The model combines the deformable convolutional network(DCN) with the original YOLOv5 to improve the sensitivity of the model to defects of different sizes and shapes. At the same time, two lightweight modules, depth-wise separable convolution(DSConv) and efficient channel attention(ECA), are introduced into YOLO model. These two modules enable the model to better understand the relationship between each channel in the input data, thus improving the detection accuracy and generalization ability of the model, and greatly reducing the complexity of the model. Then, the effectiveness of each innovation module is verified through ablation experiment and horizontal comparative experiment. Finally, a classic open source strip data set NEU-DET and an actual industrial strip data set are used to verify the effectiveness of the proposed model.

Key words: strip surface defect detection, deformable convolutional network, depthwise separable convolution, efficient channel attention, lightweight YOLOv5, image preprocessing