A Lightweight ResNet-PatchGAN Framework for Parameter-Free Multi-Noise Image Denoising

Abstract

Image denoising is a long-standing problem incomputer vision because images are often corrupted by multipletypes of noise in practical settings. This paper presents a GANbasedframework for parameter-free multi-noise removal wherethe noise type and noise parameter are not supplied to thenetwork at inference by combining established techniques: a 9-block ResNet generator for residual noise learning, a spectralnormalizedrelativistic PatchGAN discriminator for stable localtexture supervision, and a carefully balanced four-component lossconsisting of L1, relativistic adversarial, VGG perceptual, andtotal variation terms. Training is guided by a combination ofL1 reconstruction loss, relativistic adversarial loss, VGG-basedperceptual loss, and total variation (TV) loss. The dataset contains8,000 images: 2,000 clean images from the Flickr Images Datasetand 6,000 synthetic noisy images generated by applying threenoise classes—Gaussian, salt-and-pepper, and speckle—to theclean images. The clean source images and their noisy variantsare partitioned into training, validation, and test sets using a70/20/10 split. The experimental results indicate stable denoisingperformance across all three noise types with a mean PSNR of31.54 dB. A detailed ablation study evaluates the contribution ofeach loss term. This work serves as a reproducible frameworkforcombine residual learning, spectral normalization, and perceptuallosses for parameter-free restoration across the three typesof synthetic noise. Its modest computational requirements alsomake it suitable for resource-constrained applications includingedge devices and medical imaging systems.