Shallow Convolutional Neural Network for Defect Classification in Composite Materials Using Terahertz Images

Terahertz (THz) imaging has emerged as a highly promising technique for nondestructive testing (NDT), particularly for more advanced composite materials that are widely used in aerospace, automotive, and construction industries. The key strength of THz radiation lies in its ability to penetrate these materials, providing an effective means of detecting internal defects, such as cracks, voids, and delamination, without causing any damage to the material itself. However, while THz imaging is powerful, the quality of the images is often influenced by several system-dependent factors, including resolution, signal-to-noise ratio, etc. These limitations can compromise the accuracy and efficiency of defect detection. Integrating artificial intelligence (AI) algorithms, particularly convolutional neural networks (CNNs), has become crucial to overcome these challenges. AI can significantly enhance the classification process by facilitating in automatic defect detection, providing faster, more accurate inferences from the THz images. In this study, the authors have developed a simple yet effective shallow CNN for the automatic classification of defects in composites using THz images. The CNN is trained to classify defects into thirteen categories: crack, debond, delamination, groove, impact damage, indentation, void, and more. Despite the dataset being small (85 THz images) and highly imbalanced, the CNN model has been meticulously tuned by optimizing the model hyperparameters, to attain an impressive classification accuracy of over 99%. This demonstrates the capability of combining AI with THz imaging to achieve reliable, efficient, and automated defect detection in composite materials.