Soft computing-driven infrared UAV image synthesis from visible light: Leveraging multi-scale adaptation and physical consistency for real-world anti-UAV applications
Bowei Xu, Zhong Jianlin, Zhuo Changfei, Ji Wenxuan, Yang Huaxin, Luo Xiang, Ding Zhixin · Applied Soft Computing · 2026
This paper addresses three critical challenges in air-to-air infrared unmanned aerial vehicle (UAV) target detection: data scarcity, physical inconsistency, and limited multi-scale adaptability. Visible-to-infrared UAV image synthesis is inherently an uncertain cross-modal mapping task, as the relationship between visible appearance and infrared radiation is incomplete, scene-dependent, and difficult to characterize with a deterministic rule. To this end, we propose the Multi-Scale Physics-informed framework (MSPhys) , a hybrid soft-computing framework that integrates deep generative learning, adaptive feature modulation, physically informed modeling, and multi-scale discrimination for high-fidelity visible-to-infrared image synthesis. Specifically, MSPhys incorporates: 1) a physics-aware adaptive normalization layer (Phys-AdaIN + ) that embeds TeV temperature-emissivity decomposition into feature modulation; 2) a dual-path encoder with a Laplacian of Gaussian (LoG) filter to enhance multi-scale representations and reduce blur artifacts; 3) a deformable multi-scale sparse cross-attention module (DMSC) to improve cross-modal feature alignment and suppress interference; and 4) a deformable multi-scale discriminator (DMDisc) to enhance robustness in complex scenes and multi-scale UAV representation. Experiments on the Det-Fly and UAVE datasets show that MSPhys achieves consistently superior overall performance across image-quality and downstream detection evaluations compared with existing methods. These results demonstrate the effectiveness and practical value of MSPhys for infrared UAV data generation and anti-UAV perception.