Generating NIR images from RGB via Neural Style Transfer

This thesis is part of a collaborative project between the University of Pisa, SmaRTy (a European company specialized in computer systems), and Clermont Auvergne University. It focuses on patient monitoring in hospital environments using RGB and NIR (Near-Infrared) video surveillance. The main challenge addressed is the reduced performance of motion recognition models in low-light conditions typical of NIR footage. To improve recognition in such scenarios, this work explores the generation of synthetic NIR images from RGB inputs using multispectral image analysis and style transfer techniques based on Generative Adversarial Networks (GANs). The goal is to demonstrate how these models can enhance image- and video-based systems while reducing reliance on domain-specific real data. Three GAN-based methods are investigated: CycleGAN, Pix2Pix, and Pix2PixHD. Cy- cleGAN performs unpaired image-to-image translation. Pix2Pix is a conditional GAN that uses paired images for more accurate results, while Pix2PixHD, typically used in paired settings, is applied here in an unpaired context to take advantage of its high-resolution output. These models are evaluated on three public datasets and one private dataset collected with SmaRTy and Clermont Auvergne Univer- sity. While further tuning is needed, the results show promising potential. The experiments confirm the utility of GANs for realistic NIR synthesis and improving recognition in low-light conditions.