基于STM R-CNN的热轧带钢表面缺陷检测

为了提高工业热轧带钢表面缺陷检测的检测精度, 将深度学习研究领域的前沿技术应用于带钢表面缺陷检测. 提出了一种以Swin Transformer作为骨干特征提取网络, 级联多阈值结构作为输出层的热轧带钢表面缺陷检测算法. 将Transformer结构应用于带钢表面缺陷检测领域, 与单纯基于卷积网络的深度学习目标检测算法相比, 能够达到更加精确的检测效果. 首先, 使用Swin Transformer作为骨干特征提取网络代替常规的残差网络结构, 增强特征网络对隐含在图像中的深层语义信息的摄取能力. 其次设计多级联检测结构, 设置逐级的IoU阈值, 实现检测精度与阈值提升的权衡. 最后使用柔性非极大值抑制(Soft-NMS)、FP16混合精度训练和SGD优化器等训练策略加速模型收敛和提升模型性能. 实验结果表明: 本文算法在工业热轧带钢数据集(NEU-DET)上相较于YOLOv3、YOLOF、DeformDetr、SSD512和SSDLit等深度学习算法都有更好的检测效果, 在裂纹(crazing, Cr)、夹杂(inclusion, In)、斑块(patches, Pa)、麻点(pitted surface, PS)、压入氧化铁皮(rolled-inscale, RS)、以及划痕(scratches, Sc)等表面缺陷检测中训练速度和检测精度都有显著的提升, 漏检率显著降低.

Surface Defect Detection of Hot-rolled Strip Steel Based on STM R-CNN

The cutting-edge technology in deep learning is applied to surface defect detection of strip steel for the accuracy improvement in surface defect detection of industrial hot-rolled strip steel. Therefore, a surface defect detection algorithm for hot-rolled strip steel is proposed, which takes Swin Transformer as the backbone feature extraction network and cascaded multi-threshold structure as the output layer. Compared with the deep learning target detection algorithm based solely on convolutional networks, the detection algorithm using the Transformer structure can achieve more accurate detection results. Specifically, first, Swin Transformer is used as the backbone feature extraction network to replace the conventional residual network structure and thus enhance the ability of the feature network to capture the deep semantic information implicit in an image. Secondly, a multi-cascade detection structure is designed, and step-by-step IoU thresholds are set to achieve the balance between detection accuracy and threshold improvement. Finally, training strategies such as soft non-maximum suppression (Soft-NMS), FP16 mixed precision training, and SGD optimizers are employed to accelerate model convergence and improve model performance. The experimental results reveal that the proposed algorithm has better detection performance on the industrial hot-rolled strip steel data set (NEU-DET) than the deep learning algorithms such as YOLOv3, YOLOF, DeformDetr, SSD512, and SSDLit. Additionally, the training speed and detection accuracy are significantly improved in the surface defect detection of crazing (Cr), inclusion (In), patches (Pa), pitted surface (PS), polled-in scales (RS), scratches (Sc), and other surface defects, and the missed detection rate is greatly reduced.