Heat transfer model-based approaches to mitigate performance loss in thermal infrared face recognition under cross-thermal conditions

Thermal infrared face recognition is an important approach for identity verification in security and surveillance applications. However, in real-world scenarios, thermal conditions are inherently variable, and such variations can lead to significant changes in facial temperature distributions, resulting in severe degradation in recognition performance. Although deep learning has shown strong capabilities in recognition tasks, collecting thermal infrared images under diverse thermal conditions remains challenging, and limited homogeneous data fundamentally restricts model performance. This work aims to address these challenges in thermal vision from a physical heat transfer perspective. A heat transfer model is first constructed, followed by the inversion of facial thermal parameters and subsequent data simulation, enabling the generation of additional data under unseen conditions to enrich the data set and improve recognition accuracy. To this end, a facial heat transfer model adapted to thermal infrared images is proposed for thermal parameter estimation and forward simulation. In addition, a landmarking strategy specifically tailored to the thermal characteristics of facial tissues is introduced to support accurate facial alignment,data statistics and simulation. Experimental results demonstrate that the proposed method has strong capability in simulating facial temperature distributions and significantly improves recognition performance under cross-thermal conditions, achieving Rank-1 recognition improvements of more than 10% in challenging scenarios. This approach not only addresses the challenge of insufficient training caused by limited heterogeneous thermal data, but also establishes a connection between heat transfer theory and thermal infrared face recognition, enabling effective cross-domain application.