基于深度学习的激光剪切散斑干涉无损检测缺陷识别

剪切散斑干涉技术通过测量物体表面变形来推断其内部缺陷，具有高灵敏度、检测范围广、精度高等优点，是一种极具潜力的复合材料无损检测技术。目前缺陷识别主要采用人工方式，而人工识别不但检测效率低且受到专业性限制。为了提高剪切散斑干涉无损检测方法中的缺陷识别精度和效率，本文提出基于深度学习剪切散斑干涉缺陷识别方法。利用高精度四步相移技术获取剪切散斑相位条纹高质量成像；引入了应用广泛的YOLOv5和Faster R-CNN目标检测算法，通过实验采集了大量的缺陷图像，分别用YOLOv5和Faster R-CNN两种算法获得训练模型。然后将这两种模型分别应用于剪切散斑干涉无损检测中的复合材料缺陷检测。最后，实验从检测速率和检测精度方面对模型识别效果进行了对比分析。实验结果表明，激光剪切散斑干涉技术结合深度学习的方法能有效地实现剪切散斑干涉无损检测的缺陷自动识别，Faster R-CNN和YOLOv5的检测速率分别能达到11 f/s和50 f/s,并且两种深度学习算法的平均精度均能达到92%以上，验证了提出方法的可行性。

Defect recognition of laser shearographic nondestructivetesting based on deep learning

Shearography deduce its internal defects by measuring the surface deformation of the object,which has has the advantages of high sensitivity,wide detection range and high accuracy and is a potential non destructive testing technology for composite materials.At present,the defect detection is mainly done by artificial means.The artificial recognition not only has low detection efficiency,but also is limited by professionalism.To improve the accuracy and efficiency of defect recognition in shearography,a defect recognition method based on deep learning is proposed in this paper.To obtain high quality images of shearography,a high precision four step phase shifting technology is adopted in this paper.The widely used YOLOv5 and Faster R CNN target detection algorithms are introduced to achieve automatic defect detection.A large number of defect images are acquired in experiments,and the training models are obtained using YOLOv5 and Faster R CNN algorithms respectively.These two models are applied to the defect detection of composites.Finally,the recognition effect of the models is compared and analyzed from two aspects of detection efficiency and accuracy.The experimental results show that the defects of composites can be automatically recognized by combining shearography with deep learning.The defect detection can reach a rate of 11 f/s (frames per second)and 50 f/s for the Faster R CNN and YOLOv5 respectively.The average accuracy of the two deep learning algorithms can reach more than 92%,which verifies the feasibility of the proposed method.