High-quality infrared image synthesis via a three-stage cascaded generative adversarial network

This paper proposes a novel three-stage cascaded generative adversarial network (GAN) for generating high-quality infrared images from visible light images. The proposed framework comprises three key stages: a U-Net network for initial image segmentation and structural information extraction, an improved DoubleU-Net network for detail enhancement and edge contour strengthening, and a noise reduction network for denoising and final image refinement. By incorporating residual connections, atrous spatial pyramid pooling (ASPP), and squeeze-and-excitation (SE) blocks, the network effectively captures multi-scale features and refines image details. Experimental results, evaluated both qualitatively and quantitatively using metrics such as Peak Signal-to-Noise Ratio (PSNR), Structure Similarity Index Measure (SSIM), and Root Mean Square Error (RMSE), demonstrate that the proposed model outperforms existing methods in generating infrared images with richer texture details and higher fidelity. The open-source code and datasets, available at https://github.com/liangcheng1231/three-stage-GAN/tree/master , along with detailed usage guidelines, facilitate the reproduction of our findings and further research in this area.