基于KAS-YOLO的钢板表面缺陷检测

Surface Defect Detection of Steel Plate Based on KAS-YOLO

针对当前基于视觉的钢板表面缺陷检测方法对小缺陷目标识别困难造成检测精度低的问题,提出一种基于YOLOv5s的钢板表面缺陷检测模型KAS-YOLO。首先,通过使用空洞空间金字塔池化模块来获取更大的感受野以提取更多表面缺陷的特征信息,并融入坐标注意力机制来提高特征提取能力;其次,用K-means算法聚类得到更匹配的锚框,不仅增加了正样本的数量,而且加速了模型的收敛;最后,采用SIoU损失函数来进一步提升模型对表面缺陷目标的定位和检测能力。实验结果表明,提出的KAS-YOLO模型对钢板表面缺陷的检测精度和速度优于Faster R-CNN、SSD、RetinaNet和YOLOv5s等主流检测方法。

As the difficulty in identifying small defect targets,the vision-based of steel plate surface defect detection method has the problem of low detection accuracy.The model KAS-YOLO based on YOLOv5s is proposed for steel plate surface defect detection.Firstly,the hollow space pyramid pooling module is used to obtain a larger receptive field to extract more feature information of surface defects,and coordinate attention mechanism is chosen to improve the feature extraction ability.Then the K-means algorithm is used to obtain a more matched anchor frame,which not only increases the number of positive samples,but also accelerates the convergence of the model.Finally,the loss function of SIoU is used to improve the ability of locating and detecting for surface defect targets furtherly.The experimental results show that the proposed KAS-YOLO model is superior to the methods such as Faster R-CNN,SSD,RetinaNet and YOLOv5s in the detection accuracy and speed for the detection of steel plate surface defects.