Infrared Data Generation Based on the IRGAN Model

With the wide application of infrared data in aerospace, medical and other fields, there is a growing demand for high-quality infrared datasets. However, infrared datasets are relatively scarce due to the difficulty of acquiring infrared data. In contrast, visible data are easy to access, but their imaging capability is limited in specific scenes. To alleviate the problem of insufficient infrared datasets, we explore the use of deep learning techniques to convert visible data into infrared data. However, in the existing deep learning-based conversion techniques, the quality of the converted infrared data is not high when dealing with visible data containing complex multi-target semantics. To overcome this challenge, this paper innovatively proposes the Sem-IRGAN model. We introduce additional semantic segmentation data as input to the original IRGAN model. This achieves the translation of visible data and its corresponding semantic segmentation data into high-quality infrared data under a supervised learning framework. The CUS3D dataset's experimental findings indicate the efficacy of the Sem-IRGAN model. In comparison with the existing methods like Instagan and IRGAN, Sem-IRGAN outperforms them across seven key performance indicators: PSNR, SSIM, RMSE, FSIM, Cosine Similarity, LPIPS, and FID.