| Title |
Thermal Face Recognition via Thermal-to-Visible Translation with Residual-Shifting Diffusion Model |
| Authors |
홍채희(Chae-Hui Hong) ; 유훈(Hoon Yoo) |
| DOI |
https://doi.org/10.5370/KIEE.2026.75.7.1553 |
| Keywords |
Face Recognition; Thermal Face Recognition; Modality Translation; Thermal-to-Visible; Diffusion Model |
| Abstract |
This paper proposes an efficient and practical framework for thermal face recognition. The proposed method is designed as a Thermal-to-Visible (T2V) translator based on a residual-shifting diffusion model that converts thermal images into visible face images. The transformed images are then fed into a pre-trained high-performance visible-light face recognition model to extract and utilize facial embeddings. Thermal imagery contains limited visual cues for identity discrimination, and the lack of large-scale thermal training data restricts performance improvements through direct model retraining. To address this issue, the proposed method integrates modality-specific autoencoders with a residual-shifting diffusion process to develop a T2V translator optimized for modality transformation. Compared with conventional diffusion-based generative models, the proposed method significantly improves inference efficiency by performing the transformation with only 15 sampling steps. The generated visible images are input into a standard visible-light face recognition network to extract facial embeddings, which are subsequently used for final identity recognition. Experimental results on the SpeakingFaces dataset demonstrate that the proposed approach significantly improves both face detection and recognition performance when applied to a visible-light recognition pipeline, confirming that the proposed framework achieves both high inference efficiency and effective thermal face recognition. |