Infrared images are essential for low-light object recognition and scene understanding; however, their widespread use is limited by high acquisition costs. An alternative is to convert readily available visible images into infrared ones. Deep learning offers a promising solution to this problem. Nevertheless, existing methods often struggle with multi-scale targets and generate blurred boundaries. To address these issues, we propose SF-GAN, a Spatial-Frequency Joint Generative Adversarial Network for visible-to-infrared translation. SF-GAN consists of a generator and a discriminator. The generator adopts a U-Net architecture. In the encoder, we introduce a spatial pyramid pooling module with parallel spatial attention to fuse multi-scale features and emphasize key regions. In the decoder, we design an edge-guided feedforward module to jointly preserve frequency information and enforce edge constraints. In addition to the standard adversarial and L1 losses, we incorporate a perceptual loss to enhance human-perceived visual quality and a frequency-domain loss to preserve sharp boundaries. Experiments on four public datasets (KAIST, FLIR, DayDrone, and AVIID) demonstrate that SF-GAN generates infrared images with clear details and sharp boundaries, outperforming several popular translation methods (Pix2Pix, CycleGAN, PID, and DR-AVIT). Moreover, with only 20.16M parameters and 9.86ms inference time, SF-GAN achieves the highest PSNR and lowest MSE across all four datasets and remains competitive on other metrics. Furthermore, SF-GAN demonstrates competitive data augmentation performance, yielding a modest improvement in YOLOv11 detection. Our code is available at https://github.com/liutian123/SF-GAN .
Liu et al. (Thu,) studied this question.