基于改进YOLOv5模型的船舶焊缝缺陷检测方法

Ship welding seam defect detection method based on improved YOLOv5 model

当前船舶焊缝缺陷检测主要是通过人工目视的手段检查焊缝的射线探伤图像进行的,存在耗时长、工作量大、效率低的问题,为此提出了一种基于改进YOLOv5模型的船舶焊缝缺陷检测方法.首先对1 152张船舶焊缝射线图像进行标注,建立船舶焊缝射线图像数据集;然后根据船舶焊缝缺陷几何尺寸小、特征不明显的特点,对YOLOv5模型进行改进.通过对图像进行正弦灰度变换,提高缺陷处的对比度.加入卷积注意力模块(CBAM),增大感兴趣区域的权重.增加检测尺度,提高对微小目标的检测精度.计算对比检测结果表明,使用改进的YOLOv5模型对船舶焊缝缺陷进行识别,使精确度从95.3%提高到98.4%,召回率从77.5%提高到77.9%,交并比为0.5时的平均精确度从81.5%提高到84.2%,证明该方法可以有效地改进船舶焊缝缺陷检测的效果.

At present, ship welding seam defect detection still relies on manual visual checking of radiographic image, which causes long time-consuming, large workload and low efficiency. To solve this problem, a ship welding seam defect detection method is proposed based on improved YOLOv5 model. Firstly, a radiographic image dataset is established, which contains 1 152 ship welding seam radiographic images with annotation. According to the characteristics of small size and insignificant feature of ship welding seam defect, the YOLOv5 model is improved. The contrast ratio of defect area is raised by performing sine grayscale transformation of image. Convolutional block attention module (CBAM) is added to increase the weight of interested area, and the detection scale is increased to improve the detection accuracy of small targets. By calculating the comparative test results, it is shown that the identification of ship welding seam defects using the improved YOLOv5 model can improve the accuracy from 95.3% to 98.4%, the recall rate from 77.5% to 77.9%, and the average accuracy with 0.5 of intersection over union from 81.5% to 84.2%, which indicates that the proposed method can effectively improve the effect of ship welding seam defect detection.