Generative facial prior embedded degradation adaption network for heterogeneous face hallucination

In real-world long-range surveillance systems, thermal face images captured from a distance suffer from low resolution and noise, posing challenges for thermal-to-visible face image translation. Current methods assume similar resolutions and noise-free conditions between thermal and visible images, limiting their applicability. To address these issues, we propose the Generative Facial Prior Embedded Degradation Adaption Network (GDANet), which synthesizes high-quality visible images from low-quality thermal images. GDANet combines pretrained Generative Adversarial Network (GAN) blocks with a U-shaped deep neural network (DNN) to incorporate faithful facial priors, including geometry, facial textures, and colors. Additionally, an unsupervised degradation representation learning scheme is developed to capture abstract degradation representations of degraded thermal images in a representation space. This approach allows GDANet to adapt spatial features based on the degradation representation, striking a balance between fidelity and texture faithfulness using degradation-aware feature fusion (DAFF) blocks. Experimental results demonstrate that GDANet outperforms state-of-the-art methods, showing its effectiveness in handling real-world low-quality thermal images across diverse practical applications.