Abstract: The defect detection on the surface of the steel plate is the basis and key to analyze and judge the quality of the steel plate.Aiming at the problems of low detection efficiency and poor accuracy of small defects such as cracks on the surface of steel plates,a real-time detection algorithm for surface defects of steel plates integrated with Transformer was proposed.By integrating the Coordinate Attention(CA)module and the Dynamic Shift Max(DY)activation function,the CA-Bneck module was constructed to improve the representation ability of defect features.The MobileNetV3,CA-Bneck and Transformer were encoded,and the modules were integrated to build a new backbone feature extraction network MobileNetV3-CATr,which was used to reduce the complexity of the model.A BiFPN-Lite module was proposed to fuse more defect features without increasing the complexity of the model,and the defect information was output through YOLO Head.The experimental results on the open data set NEU-DET of hot rolled steel showed that the proposed algorithm achieved a balance between performance and speed.Compared with YOLOv4,the mAP value was increased by 5.96%,and the speed reached 20.1FPS,which could effectively complete the real-time and high-precision detection of steel plate surface defects.

Key words: machine vision, defect detection, MobileNetv3, Transformer block

摘要： 钢板表面的缺陷检测是分析和判断钢板质量的基础和关键。针对钢板表面龟裂等小缺陷检测效率低、精度差的问题,提出一种融合Transformer的钢板表面缺陷实时检测算法。首先,融合协调注意力(CA)模块以及最大动态转移(DY)激活函数构建CA-Bneck模块,提高缺陷特征的表示能力;其次,将MobileNetV3、CA-Bneck以及Transformer编码模块相融合,构建一种新的主干特征提取网络MobileNetV3-CATr,用于减轻模型的复杂度;最后,提出一种BiFPN-Lite模块,使得模型复杂度不增加的条件下融合更多缺陷特征;并通过YOLO Head输出缺陷的信息。在热轧钢公开数据集NEU-DET上实验结果表明,所提算法在性能和速度之间取得了平衡,mAP值相较于YOLOv4提升了5.96%,速度达到了20.1FPS,能够有效地完成钢板表面缺陷的实时和高精度检测。

关键词: 机器视觉, 缺陷检测, MobileNetv3网络, Transformer模块