Scholarlink

<p>Nondestructive testing (NDT) of composite materials presents significant challenges due to their complex, anisotropic, and heterogeneous internal structures, which often limit the efficacy of conventional inspection techniques. In this study, a terahertz (THz) imaging system is employed to address these limitations by enabling identification of subsurface and bulk defects in composite materials. However, the THz images are highly noisy with low resolution and high background fluctuations. Hence, for automated defect localization, it is crucial to employ the state-of-the-art artificial intelligent algorithms yielding in AI augmentation of the THz images. Here, a dataset comprising over 600 THz images, categorized into eighteen distinct defect classes, has been employed for automated defect localization studies. Due to the poor THz image quality, it becomes extremely challenging to localize the defects in the composites manually or visually. Therefore, it is essential to develop an efficient AI-based framework capable of localizing defects with increased accuracy. Three object detection frameworks—Region-based Convolutional Neural Network (RCNN), YOLOv8l, and Faster R-CNN — have been evaluated in terms of their detection accuracy, localization precision, and computational efficiency. The RCNN model achieved a maximum precision and recall of 0.9738 and 0.9737, respectively, while YOLOv8l attained a precision and recall of 0.9550 and 0.9459, respectively. Faster R-CNN attained perfect classification scores, with precision and recall of 1.0000, and outperformed other models in terms of inference speed, requiring only 1.2&#xa0;ms per image, which has not been reported elsewhere. Given its superior balance of classification performance, localization, and real-time inference capability, Faster R-CNN has been identified as the most effective model for THz-based defect detection in composites. However, YOLOv8l demonstrated superior localization accuracy with the highest mean IoU of 0.9872. These findings demonstrate the potential of AI augmented THz imaging in enabling robust, automated, and scalable NDT solutions for advanced composite inspections.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Defect Localization in Composites Using AI-Augmented Thz Imaging

  • 期刊论文

  • 发表日期：

    2026年3月29日

查看全文
  • Battipally Jeevakranthi Goud,
  • Bobba Krishna Sai,
  • M. Muth Tamil Pooja,
  • S. Vinson Joshua,
  • A. Mercy Latha

摘要

Nondestructive testing (NDT) of composite materials presents significant challenges due to their complex, anisotropic, and heterogeneous internal structures, which often limit the efficacy of conventional inspection techniques. In this study, a terahertz (THz) imaging system is employed to address these limitations by enabling identification of subsurface and bulk defects in composite materials. However, the THz images are highly noisy with low resolution and high background fluctuations. Hence, for automated defect localization, it is crucial to employ the state-of-the-art artificial intelligent algorithms yielding in AI augmentation of the THz images. Here, a dataset comprising over 600 THz images, categorized into eighteen distinct defect classes, has been employed for automated defect localization studies. Due to the poor THz image quality, it becomes extremely challenging to localize the defects in the composites manually or visually. Therefore, it is essential to develop an efficient AI-based framework capable of localizing defects with increased accuracy. Three object detection frameworks—Region-based Convolutional Neural Network (RCNN), YOLOv8l, and Faster R-CNN — have been evaluated in terms of their detection accuracy, localization precision, and computational efficiency. The RCNN model achieved a maximum precision and recall of 0.9738 and 0.9737, respectively, while YOLOv8l attained a precision and recall of 0.9550 and 0.9459, respectively. Faster R-CNN attained perfect classification scores, with precision and recall of 1.0000, and outperformed other models in terms of inference speed, requiring only 1.2 ms per image, which has not been reported elsewhere. Given its superior balance of classification performance, localization, and real-time inference capability, Faster R-CNN has been identified as the most effective model for THz-based defect detection in composites. However, YOLOv8l demonstrated superior localization accuracy with the highest mean IoU of 0.9872. These findings demonstrate the potential of AI augmented THz imaging in enabling robust, automated, and scalable NDT solutions for advanced composite inspections.