【作者】 刘凯；

【导师】 郭文明；

【作者基本信息】 北京邮电大学 ， 软件工程， 2020， 硕士

【摘要】 无损检测技术是焊接质量检测的一种重要的手段,并应用于航空航天、建筑、管道运输等领域。焊接过程中产生的缺陷对设备的使用质量和使用寿命有至关重要的影响。检测焊接部分是否有缺陷存在,不仅可以验证被测目标焊缝质量是否达到焊接标准,还能及时发现重大缺陷,及时修复,提高焊缝质量。传统、常用的检测方式是使用专门的软件设备对X-射线图片进行对比度、灰度变换后人为检测,此过程不仅效率低下,而且受人为影响较大。针对于此,探索一种能自动定位、分类焊接图片中潜在的缺陷的方法极为重要且具有实际意义。通过了解焊接缺陷检测的国内外研究现状,以及当前深度学习的发展,尤其是目标检测领域的研究进展,确定了本文的研究方向。借助深度学习领域Faster RCNN模型这一经典的卷积神经网络模型,简化原来常用的焊接缺陷检测的步骤,实现焊接缺陷的全自动检测。在数据集预处理部分,需要对原图进行了焊接区域的提取,对此本文提出了一种基于单位面积平均对比度值和平均灰度值的方法来锁定焊接区域的边界,编写标注程序生成VOC数据集格式。利用自适应的直方图均衡化和两次中值滤波对提取焊接区域后的图片进行图像增强和去噪处理。在模型训练过程中,本文尝试了特征增强、标签平滑等方法对模型的优化,以提高模型的检测效果。本文最终根据研究结果实现了一款能自动检测缺陷类型以及缺陷区域坐标的软件,这是与传统方式完全不同的尝试,深度学习与工业中的具体应用相结合,让模型从数据集中训练和学习,来达到自动检测的效果,这将会是今后无损检测领域研究的一个热门方向和趋势。更多还原

【Abstract】 Non-destructive testing(NDT)technology is an important method for welding quality inspection and is used in aerospace,construction,pipeline transportation and other fields.The defects produced in the welding process have a crucial impact on the quality and service life of the equipment.Detecting whether there are any defects in the welding object,can not only verify whether the quality of the weld of the tested target has reached the welding standards,but also find destructive defects in time to repair them and improve the quality of the welding.The traditional and commonly method of welding detection is to artificially detect the X-ray images after the contrast and grayscale transformation using special software equipment by the NDT inspectors.This process is not only inefficient,but also greatly affected by human factors.In view of this,it is extremely important and practical to explore a method that can automatically locate and classify potential welding defects in welding images.By understanding the current research status of welding defect detection at home and abroad,and the current development of deep learning,especially the research progress in the field of object detection,the research direction in this paper is determined.With the help of the classic convolutional neural network model in the field of deep learning,the Faster RCNN model,the steps of the original welding defect detection are simplified and the automatic detection of welding defects is researched.The images in the data set used in the experiment are digital pictures obtained from the actual X-ray film collected from various places and converted by a special film scanning equipment.The adaptive histogram equalization and two median blur were used to perform image enhancement and denoising on the extracted welding area.In the model training process,this paper attempts to optimize the model by methods such as feature enhancement,label smoothing and other tricks to improve the detection performance.Based on the research results,this paper finally realized a system that can automatically detect the types of defect in the picture and the coordinates of the defect area.This is a completely different attempt from the traditional methods.Deep learning is combined with specific applications in industry,allowing the model to train and learn from the data set to achieve the performance of automatic detection.This will be hot directions and trends in the field of non-destructive testing in the future.更多还原