ABSTRACT

Continuous casting is a dominant process for steel production with high productivity, low cost and high automation, but it always suffers from corner defects in the strand. Thus, an in-situ and highly efficient detection of the strand corner crack is very urgent for high-quality steel production. In the present study, several models, namely, YOLOv5x, YOLOv5-S (YOLOv5 + ShuffleNet v2), YOLOv5-SF (YOLOv5 + ShuffleNet v2 + Focus) and YOLOv5-SFA (YOLOv5 + ShuffleNet v2 + Focus + Adam optimizer), are proposed. The experimental results show that among the four models, the mAP for YOLOv5-SFA increases fastest and the number of epochs for mAP reaches the maximum is least. The loss value is less than 0.01 and the training time is 0.369 h, which is reduced by 58.86% with the comparison of YOLOv5x. When only 100 images are used as training data, the detection accuracy is 99.64%, which increases 11.19% with comparison of YOLOv5x, and the detection time is only 0.021 s.

摘要

连铸是生产率高、成本低、自动化程度高的钢铁生产的主导工艺，但铸流中常存在角部缺陷。因此，对钢绞线转角裂纹进行原位高效检测成为优质钢材生产的当务之急。在本研究中，几个模型，即 YOLOv5x、YOLOv5-S（YOLOv5 + ShuffleNet v2）、YOLOv5-SF（YOLOv5 + ShuffleNet v2 + Focus）和 YOLOv5-SFA（YOLOv5 + ShuffleNet v2 + Focus + Adam 优化器），是建议的。实验结果表明，在四个模型中，YOLOv5-SFA的mAP增长最快，mAP达到最大值的epoch数最少。损失值小于0.01，训练时间为0.369 h，与YOLOv5x相比减少了58.86%。当仅使用 100 张图像作为训练数据时，