Visible-to-Infrared Image Translation via Disentangled Imaging Attribute Transfer

Visible (VI)-to-infrared (IR) image translation aims to synthesize IR images from easily accessible VI acquisitions, offering a cost-effective alternative to expensive IR data collection in remote sensing scenarios. Despite remarkable advances achieved, existing methods still struggle to generate high-fidelity IR images. This limitation stems from their reliance on holistic image-level domain-to-domain mapping, which prioritizes global visual alignment while failing to bridge the gap between the imaging attributes of the two modalities. To tackle this critical issue, we propose a novel disentangled architecture for VI-to-IR image translation, which abandons direct global domain mapping and instead explicitly models the transfer of thermal imaging attributes. Specifically, we first decompose the latent representation of each image into two disentangled components: modality-invariant semantic content and modality-specific imaging attributes. Subsequently, an attribute transfer module is designed to map VI-specific imaging attributes to the IR domain. Finally, the transferred IR attributes and preserved semantic content are fed into a conditional generative adversarial network (cGAN) to synthesize IR images with authentic thermal properties. Furthermore, we adopt a multi-stage end-to-end training paradigm integrated with dedicated hybrid loss functions to ensure stable model convergence and enhance the fidelity of the generated IR images. Extensive experiments conducted on the AVIID, Day-DroneVehicle, and Night-DroneVehicle benchmarks demonstrate that our method outperforms multiple state-of-the-art (SOTA) methods in both quantitative and qualitative evaluations. The source code is available at https://github.com/silver-hzh/VIAT.