Infrared Image Generation Based on Improved CycleGAN

In the field of image processing, visible images, although commonly used, have limited performance in low light or bad weather. Infrared images provide an effective solution with their superior performance in dark light and complex weather conditions. However, traditional infrared images are expensive to acquire and have poor environmental adaptability. To overcome these limitations, an improved unsupervised image transformation model, IF-CycleGAN is advocated in this paper. This model incorporates a self-attention subsystem and a feature fusion strategy on the basis of CycleGAN, which significantly improves the details and quality of infrared image generation. Through experimental evaluation on the AVIID dataset, IF-CycleGAN performs well in image quality evaluation metrics. In terms of PSNR, IF-CycleGAN improves 1.22 points to 19.01 points compared to CycleGAN, showing a higher match with real images. SSIM also improves to 0.66, which proves the advantages of IF-CycleGAN in terms of image structure and texture realism. In addition, the LPIPS is reduced to 0.23, which further indicates that the image generated by IF-CycleGAN is visually closer to the real infrared image. This verifies the effectiveness of IF-CycleGAN in image transformation tasks, where IF-CycleGAN shows its advantages in visual content retention, structural characterization and infrared properties.