Unsupervised clustering of pulsed infrared thermography for automated assessment of a Qing Dynasty brass mirror

Non-destructive assessment of complex multi-layered structures in cultural heritage artifacts, such as Qing Dynasty lacquer-coated brass mirrors, is critical for conservation and technical study. Pulsed thermography (PT), an active infrared thermographic technique, provides pixel-wise temporal thermal responses sensitive to subsurface variations, but machine-learning-based interpretation is often limited by the scarcity of annotated data. To address this challenge, a fully unsupervised framework combining Thermographic Signal Reconstruction (TSR), a Deep Autoencoder (DAE), and a Gaussian Mixture Model (GMM) is proposed to cluster thermal responses and extract thermally distinct subsurface features without prior labeling. The method was validated on a custom-designed test sample and successfully applied to a Qing Dynasty lacquered brass mirror, revealing spatially distinct thermal-response patterns associated with the lacquer-adhesive coating system. This label-free approach provides a practical tool for non-destructive evaluation of cultural heritage objects, offers new insights into historical craftsmanship, and is extendable to other metal-coating systems, supporting preservation strategies and informed conservation decision-making.