基于机器学习的连铸坯低倍缺陷检测

针对连铸坯低倍缺陷评级问题,建立了一种基于深度学习框架的系统解决方案。基于显著目标提取的深度网络模型进行了连铸坯区域提取和几何校正;基于YOLO V4的缺陷目标检测算法,进行了检测类缺陷的检测与识别,以标准的平均正确率AP(Average Precision)指标作为评价指标,“中心缩孔”、“中心疏松”、“非金属夹杂”、“皮下气泡”和“中心偏析”五类缺陷检测的AP分别达到了82.19%、97.63%、54.27%、66.20%和29.29%;基于MASK RCNN的缺陷实例分割算法,进行了分割类缺陷的检测与识别,以标准的AP(0.5-0.95)作为评价指标,“中心裂纹”、“角部裂纹”、“中间裂纹”和“皮下裂纹”4类缺陷检测和分割的AP(0.5-0.95)达到了0.78,特别地,以生产应用角度出发,AP(0.5)达到了0.96,可以较好地满足缺陷检测需要。

Defect detection of billet macrostructure based on machine learning

Aiming at the problem of low magnification defect rating of continuous casting billets, a system solution based on deep learning framework was established. Based on the defect target detection algorithm of YOLO V4, the detection and recognition of defects of detection class are carried out. The standard Average Precision (AP) index is used as the evaluation index. The AP of “central pipe”, “central porosity”, “nonmetallic inclusion”, “subsurface blowhole” and “central segregation” reached 82.19%, 97.63%, 54.27%, 66.20% and 29.29% respectively. The defect instance segmentation algorithm based on MASK RCNN was used to detect and identify segmented defects. Taking the standard AP(0.5-0.95) as the evaluation index, the AP(0.5-0.95) for detecting and segmentation of four types of defects, namely, “central crack”, “corner crack”, “middle crack” and “subcutaneous crack”, reached 0.78. In particular, From the perspective of production and application, AP(0.5) reaches 0.96, which can better meet the needs of defect detection.