Categorized visible-to-infrared image generation based on physical properties of different targets

The mapping relationship between visible images and infrared images of different targets varies due to differences in their physical properties, such as surface emissivity, reflectivity, temperature, ambient radiation intensity, and heat dissipation effects. When using Generative Adversarial Networks for visible-to-infrared image translation, we focus not only on the overall quality of the image but also on the accuracy of target feature translation. In previous research, the authors used the AVIID-3 dataset to classify cars targets into four categories: light-colored moving cars A, dark-colored moving cars B, light-colored parked cars C, and dark-colored parked cars D. The scenes are divided into two categories: moving scenes with only moving car targets and parking lot scenes with only parked car targets. An optimal generation strategy for the AVIID-3 dataset based on Pix2pix and CycleGAN has been proposed. Although this research has made significant progress, some limitations still exist. In the AVIID-3 dataset, moving and parked car targets do not appear simultaneously in the same scene, making it impossible to evaluate different generation strategies in complex scenes. To address this issue, this study proposes a general visible-to-infrared image generation strategy for car targets. Additionally, data from complex scenes captured by drones were used to reconstruct the dataset. This approach validates that the proposed strategy is effective not only for simple scenes containing only one type of target but also for mixed scenes with random combinations of multiple targets, demonstrating its practical applicability in real engineering scenarios.