VTOI-GAN: An Enhanced Pix2Pix Model for Infrared Image Generation

Infrared imaging techniques are capable of providing reliable visual information across varying lighting conditions, particularly in environments with low or no illumination. However, traditional infrared image generation methods still face limitations in scene adaptability, image quality, and efficiency. To address this, this paper proposes VTOI-GAN, a lightweight method based on an Enhanced Pix2Pix model. Firstly, a residual dense Unet generator with multiscale feature fusion is constructed to enhance the extraction and reconstruction of image features. Secondly, a dual discriminator incorporating the Sobel operator is designed to fully exploit the high-frequency information in infrared images. Finally, a composite loss function is designed by combining adversarial loss, L1 loss, edge loss, and shallow feature loss to effectively constrain the texture and grayscale information of the infrared images generated. Compared to Pix2Pix, the infrared images generated by VTOI-GAN on the M3FD low-light dataset exhibit improvements of 11.32% in PSNR, 15.53% in SSIM, 35.17% in LPIPS and 23.33% in FID. The experimental results demonstrate that the VTOI-GAN can generate infrared images with higher fidelity, and they are visually closer to real infrared images in terms of similarity metrics.