Cross-domain generation of visible-infrared aerial images based on distribution consistency enhancement

Le résumé fourni par la source

Aerial infrared images are widely applied in the military field with excellent visibility under low-light and smoky conditions. They effectively compensate for the limitations of visible light in adverse illumination environments and can provide more clear target information. However, existing methods and models suffer from the limitation of insufficient distribution consistency between generated results and real aerial infrared images under variable illumination and complex scenes. To improve the matching degree between cross-domain generation results and real infrared distribution, and to provide effective and realistic infrared data for detection, recognition and anti-recognition tasks, a cross-domain generation method for visible-infrared aerial images based on distribution consistency enhancement is proposed. This method extracts the unique distribution features from infrared images step by step through a three-stage infrared feature extraction network: the first stage reconstructs basic features, the second stage purifies the infrared-specific distribution, and the third stage achieves infrared feature alignment. On the basis of infrared features, a bidirectional constraint mechanism for forward generation and backward reconstruction is constructed by combining a spatiotemporal infrared enhancement network with an enhanced infrared reconstruction auxiliary loss, which realizes closed-loop guidance from forward generation to backward reconstruction and enhances the model's adaptability and robustness under different illumination conditions. To further strengthen the model's ability to learn the core of infrared distribution, a pre-trained reconstruction framework is introduced. With the help of pre-trained model parameters, stronger distribution priors and structural constraints are provided for the generation process, thereby improving the model's learning effect on infrared distribution features. Experimental results show that on the DroneVehicle dataset, the Fréchet Inception Distance (FID), a distribution consistency metric, of the proposed method reaches 17.37 in complex scenes, which is significantly superior to that of existing methods, and the FID remains optimal even under complex illumination conditions such as nighttime and low light. On the AVIID dataset, the Peak Signal-to-Noise Ratio (PSNR) reaches 23.56, the Structural Similarity Index Measure (SSIM) hits 94.93, and the Learned Perceptual Image Patch Similarity (LPIPS) achieves 12.12.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.