2026 ASEE Annual Conference & Exposition

Cross-domain face recognition aims to match probe images captured in the thermal spectrum to a gallery of visible face images. It is a critical technique for night-time surveillance and security applications. However, the large modality gap between faces captured in the thermal and visible spectra makes face recognition a challenging problem. In this paper, we propose a method for thermal to visible face recognition based on coupled independent component analysis (Coupled ICA). It has been reported that independent component analysis (ICA) of natural scene patches produces a set of visual filters that resemble the receptive fields of simple cells in visual cortex and the projection matrix form a basis of images. Aiming to learn a common latent space for cross-spectral images, we propose to learn a separate set of ICA filters which represent the respective imaging system in each domain using a coupled architecture. The coupled ICA assumes the image sources in one domain to be identical to those observed in the other domain. Pairs of image patches in the two domains jointly update the projection matrix in the Coupled ICA model. The ICA filters obtained are used to transform images into a domain-independent latent space via patch-wise synthesis. The results show that the proposed method fuses the thermal and visible images and outperforms the state-of-the-art methods of cross-spectral face recognition.

In addition, the research initiated an educational component that computer science undergraduate students conducted thermal-to-visible face recognition using traditional methods including deep learning techniques. Their research findings have been presented at national and regional conferences and garnered positive feedback from the academic community.