Enhancing Image Quality by Optimizing and Fine-Tuning Multi-Fidelity Generative Adversarial Networks

Achieving a balance between image quality and computational efficiency is one of the most challenging tasks in image enhancement. Conventional single-fidelity methods focus on structural integrity; yet, they are ineffective in improving perceptual quality, resulting in unrealistic images. The limitations identified in this paper are addressed by introducing a unique Multi-Fidelity Generative Adversarial Network (MF-GAN). The quality of the image is enhanced by the dynamic integration of low-fidelity and high-fidelity models using a composite loss function and dual-generator design. This method achieves high superior perceptual quality and structural integrity while preserving computational efficiency by gradually improving image details. The model was tested on standard natural image benchmarks, including Set5, Set14, and DIV2K, it achieved a Natural Image Quality Evaluator (NIQE) of 4.55, a Peak Signal-to-Noise Ratio (PSNR) of 30.5 dB, and a Structural Similarity Index (SSIM) of 0.92, outperforming existing methods in terms of fidelity and perceptual quality when compared to techniques like ESRGAN (Enhanced Super-Resolution GAN) and RDN (Residual Dense Network). Regarding to the computational efficiency, MF-GAN has 27.6M parameters which is competitive to other models. It achieves an inference time of 1.32 seconds with 203.7G FLOPs (Floating Point Operations). Although MF-GAN is a little faster than RDN, its computational cost is higher than models like ESRGAN and SRGAN (Super-resolution GAN). These results show that MF-GAN is a promising method for image enhancement, since it effectively balances computational efficiency and perceptual image quality.