Thermal Face Recognition Based on Multi-scale Image Synthesis

We present a transformation-based method to achieve thermal face recognition. Given a thermal face, the proposed model transforms the input to a synthesized visible face, which is then used as a probe to compare with visible faces in the database. This transformation model is built on the basis of a generative adversarial network, mainly with the ideas of multi-scale discrimination and various loss functions like feature embedding, identity preservation, and facial landmark-guided texture synthesis. The evaluation results show that the proposed method outperforms the state of the art.