Multi-descriptor hybrid fusion of deep, latent and handcrafted features for reliable thermal tomographic image classification

Thermal tomographic imaging leverages boundary-level sensor data to reconstruct internal temperature fields in thermally driven flow systems, offering a non-invasive solution for industrial condition monitoring, fault prediction, and quality control. However, reliably classifying thermal tomographic images is still difficult, primarily because of intricate heating patterns and the scarcity of well-labelled datasets. In this article, a tomographic dataset is synthetically constructed using COMSOL Multiphysics with FEM-based simulations of a contour inside a hot airflow chamber, emulating boundary sensor measurements under diverse heating and airflow conditions. A multi-descriptor hybrid feature-fusion framework is proposed, integrating discriminative features from ‘ResNet-18’ backbone, latent representations using a stacked autoencoder, and handcrafted histogram of oriented gradient features refined through principal component analysis. The fused-feature representation is evaluated using linear classifiers. Experimental results demonstrate superior performance compared to individual and dual-branch models, achieving maximum classification accuracies of up to 99.05% with excellent macro F1-score, and Cohen’s kappa. Additional analyses further reveal improved feature separability and robustness under measurement noise. This framework suggests a reliable solution for automated categorisation of heating patterns in thermal tomographic imaging, with strong potential for deployment in industrial condition monitoring and non-destructive evaluation systems.