基于改进SOLOv2的转炉炼钢钢液检测

Steel liquid detection in converter steelmaking by using improved SOLOv2

转炉炼钢是冶金工业生产中的重要工艺过程,在转炉炼钢中连续实时检测钢液有助于提升炼钢的出钢产品质量和冶炼效率,针对转炉炼钢过程中炉口采集的图像相似性高、背景干扰大导致难以检测钢液的问题,提出了一种基于深度学习的实例分割方法。在SOLOv2的骨干网络中加入通道注意力机制,并对输入的图像进行去雾的预处理操作,从而获得具有高准确率的转炉炼钢钢液检测在线模型SA-SOLOv2。在转炉炼钢测试集上进行测试,结果表明,与原SOLOv2模型相比,该模型的交并比(IoU)提升了1.4%,漏检率降低了9.8%,且改进的SA-SOLOv2模型的检测时间明显少于其他基于深度学习的检测网络,提升了检测效率。

Converter steelmaking is an important process in metallurgical industry production.Continuous real-time detection of liquid in converter steelmaking helps to improve the quality of steel tapping products and smelting efficiency.To address the problem of high similarity and background interference in converter steelmaking images,an instance segmentation model based on deep learning was proposed.A channel attention mechanism was added into the backbone network of SOLOv2(Segmenting Objects by LOcations version 2),and a pre-processing operation of defogging the input images was performed to obtain an online model SA-SOLOv2(Split-Attention SOLOv2 Network)with high accuracy for converter steelmaking steel liquid detection.Results on the test set of converter steelmaking images show that compared with the original SOLOv2,the IoU(Intersection over Union)of the proposed model is improved by 1.4%and the false detection rate is decreased by 9.8%,and the detection speed of the improved SOLOv2 model is higher than other deep learning-based detection networks,which proves a significant increase in the detection accuracy and efficiency.