Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (11): 250-259.DOI: 10.3778/j.issn.1002-8331.2110-0375

• Engineering and Applications • Previous Articles     Next Articles

Micro-Defect Detection Algorithm with Improved Attention Mechanism in Section Steel Surface

YU Haitao, LI Fulong, LIU Yajiao, WANG Jiang, YU Lifeng, ZHANG Chunhui, LIU Baoshun, MA Yongfu   

  1. 1.School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China
    2.Hebei Jinxi Iron and Steel Group, Tangshan, Hebei 063000, China
  • Online:2022-06-01 Published:2022-06-01

改进注意力机制的型钢表面微小缺陷检测方法

于海涛,李福龙,刘亚姣,王江,于利峰,张春晖,刘宝顺,马永福   

  1. 1.天津大学 电气自动化与信息工程学院,天津 300072
    2.河北津西钢铁集团股份有限公司,河北 唐山 063000

Abstract: A novel surface defect detection method named DMSA-YOLOv3 based on dual multi-scale attention mechanism is proposed in order to solve the problems of low efficiency and poor accuracy of surface defects detection caused by various and numerous micro-scale defects, which achieves a rapid and accurate detection of section steel surfaces. Firstly, a dual multi-scale attention, i. e. DMSA model is constructed based on channel and space attention to screen the fuse different scale features for strengthening the weight of small-scale defect features. Secondly, the feature extraction backbone network of YOLOv3 model is processed by deep separable convolution to reduce the size of model for speeding the detection procedure, and a multi-scale long-distance context feature extraction layer in order to improve the ability of extracting small defect targets’ feature of backbone network is built. Finally, a DMSA-YOLOv3 defect detection model combined DMSA model and improved YOLOv3 model is constructed and applied to multi-scale defect detection of steel surface. The experimental results show that DMSA-YOLOV3 model has a multi-category average detection accuracy of 97.6% and a detection speed of 55.3?frame/s, which is 4.7 percentage points and 24.5?frame/s higher than YOLOv3 model respectively. The minimum size defects of 10×10(mm2) can be detected, which is 6.25 times higher than that of YOLOv3 model. The DMSA-YOLOv3 model has boosted the accuracy and speed of surface defects detection of steel prominently.

Key words: section steel, defect detection, attention mechanism, YOLOv3 model

摘要: 针对型钢表面缺陷种类多样、微小缺陷占比较大导致的检测效率低、检测精度差的问题,提出了一种基于双重多尺度注意力机制的表面缺陷检测方法DMSA-YOLOv3,实现型钢表面多尺度缺陷快速精确检测。构建了基于通道和空间的双重多尺度注意力模型DMSA,对不同尺度特征进行筛选融合,强化小尺度缺陷的特征权重;改进了YOLOv3模型,使用深度可分离卷积对DarkNet53特征提取主干网络实现轻量化处理,提高检测速度,并构建多尺度长距离上下文特征提取层,使用4种不同扩张率的并行空洞卷积替代全局池化,提高模型对小尺寸缺陷的特征提取能力;构建了融合DMSA模型和改进YOLOv3模型的DMSA-YOLOv3缺陷检测模型,并应用于型钢表面多尺度缺陷检测。实验结果表明:DMSA-YOLOv3模型具有97.6%的多类别平均检测精度和55.3?frame/s的检测速度,与YOLOv3模型相比分别提升了4.7个百分点和24.5?frame/s;最小可检出20×20像素(约10×10?mm2)缺陷,与YOLOv3模型相比提高了6.25倍,有效提升了型钢表面缺陷的检测精度与检测速度。

关键词: 型钢, 缺陷检测, 注意力机制, YOLOv3模型