Deep learning technique, as an end-to-end method, has emerged great potentials in terahertz nondestructive testing (THz NDT) due to its powerful feature extraction ability. However, traditional models rely on excessive THz data with labels, which limits the practical application of deep learning in THz NDT. To address this issue, this study proposes a novel unsupervised learning framework for THz high-resolution imaging of composite internal defects. The framework first implements the alignment operation for the acquired THz signals from sample. Then, the unsupervised learning framework, including the clustering and classification process, is proposed to identify defects. In the clustering process, a stacked autoencoder and K - Means++ algorithm are applied to cluster the input THz signals, and the clustering results are used to generate a pseudo-label dataset for subsequent classification process. In the classification process, the Long Short-Term Memory (LSTM) network is used to classify the THz dataset. In this case, the THz high-resolution imaging process can be performed based on the classification results. Finally, a series of experiments validate the effectiveness of proposed framework under little unlabeled THz data, which provides a novel solution for high-resolution imaging of composite internal defects in THz NDT.

Terahertz nondestructive testing, composite defects, unsupervised learning, high-resolution imaging