IVT-GSCNet: Structure-Consistent and Semantically Aligned Unpaired Infrared-to-Visible Translation for UAV Remote Sensing Imagery

Unmanned Aerial Vehicle (UAV) infrared remote sensing is pivotal for nighttime perception. However, infrared (IR) images inherently suffer from low resolution and a lack of chromatic information, which restricts semantic scene understanding. To address the triad of challenges—noise contamination, semantic misalignment, and structural distortion—in unpaired infrared-to-visible translation, this paper proposes IVT-GSCNet (Infrared-to-Visible Translation Gated Spatial-Channel Network), a novel framework tailored for UAV remote sensing. Specifically, a Gated Feature Refinement Module (GFRM) is introduced to adaptively suppress sensor noise and background clutter within U-Net skip connections, ensuring clean feature propagation for reconstruction. Furthermore, to achieve precise spatial alignment without paired data, we propose a Layer-weighted Contrastive Learning mechanism. By assigning higher weights to deep layers that encode global semantic features, our method significantly improves the preservation of overall scene structure and the localization of small objects in UAV imagery. Finally, a composite structural constraint—integrating edge, structural similarity, and smoothness losses—is implemented to safeguard geometric fidelity. Extensive experiments across four diverse multi-modal datasets (including primary UAV-captured sets, MSRS, and M3FD) demonstrate that IVT-GSCNet significantly outperforms state-of-the-art methods in multi-dimensional perceptual quality (e.g., LPIPS, KID, and BRISQUE) and downstream semantic segmentation. Systematic ablation studies and parameter analyses further validate the individual module contributions and confirm our model’s unique capability in balancing structural coherence and photorealism, establishing its practical value for low-illumination UAV perception.