Enhancing Perceptual Quality of Images using Deep Residual U-Net and PatchGAN Discriminator

This paper presents a novel approach for perceptually aware image enhancement utilizing a deep residual U-Net generator combined with a Patch-based Generative Adversarial Network (PatchGAN) discriminator. Our method incorporates a perceptual loss function based on a pre-trained VGG16 network, ensuring that enhanced images exhibit improved color sharpness, clarity, and retention of critical high-level features. Experiments show improved structural similarity, with SSIM of 0.9270 and FSIM of 0.9998. These results indicate both perceptual and structural enhancement, while the model efficiently generates high-quality images with minimal training. Our method consistently outperforms conventional approaches in visual appeal and computational efficiency, making it relevant for real-world applications such as radiology, self-driving vehicles, and low-light photography. Our implementation is open-sourced and available here: https://rb.gy/bp8zsl.