摘要
本文旨在利用深度学习图像处理技术,提出一种基于YOLOv5网络结构的缺陷检测算法对超声B扫描图像进行缺陷识别实验研究。通过对工件进行线性阵列相控阵超声检测,连续采集到多幅B扫描图像或实时采集视频数据,并对图像进行预处理和增强操作。在此基础上,构建神经网络模型,并对预处理后的大量数据集进行标定。使用Pytorch深度学习框架进行数据集训练,结果表明改进的YOLOv5算法对缺陷的目标检测召回率识别达到96.2%,对底波缺失的目标检测召回率达到96.3%。实验结果表明,基于改进的YOLOv5算法在mAP0.5指标上有1.4%的提升,在mAP0.5:0.95指标上有6%的检测性能提升。
This article aims to use deep learning image processing technology to propose a defect detection algorithm based on YOLOv5 network structure to conduct experimental research on defect identification in ultrasonic B-scan images.By performing linear array phased array ultrasonic testing on the workpiece,multiple B-scan images are continuously collected or video data is collected in real time,and the images are preprocessed and enhanced.On this basis,a neural network model is constructed and a large number of preprocessed data sets are calibrated.Using the Pytorch deep learning framework for data set training,the results show that the improved YOLOv5 algorithm has a recall rate of 96.2% for defective target detection,and a recall rate of 96.3% for target detection with missing bottom waves.Experimental results show that the improved YOLOv5 algorithm has a 1.4% improvement in the mAP0.5 index and a 6% improvement in detection performance in the mAP0.5:0.95 index.
作者
冯玮
FENG Wei(Department of Mechanics&Acoustics,AVIC Changcheng Institute of Metrology&Measurment,Beijing 100095,China)
出处
《无损探伤》
2024年第4期21-26,共6页
Nondestructive Testing Technology
基金
基础加强计划技术领域基金(2021-JCJQ-JJ-1269)。
