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

Abstract

Image denoising is a long-standing problem in
computer vision because images are often corrupted by multiple
types of noise in practical settings. This paper presents a GANbased
framework for parameter-free multi-noise removal where
the noise type and noise parameter are not supplied to the
network at inference by combining established techniques: a 9-
block ResNet generator for residual noise learning, a spectralnormalized
relativistic PatchGAN discriminator for stable local
texture supervision, and a carefully balanced four-component loss
consisting of L1, relativistic adversarial, VGG perceptual, and
total variation terms. Training is guided by a combination of
L1 reconstruction loss, relativistic adversarial loss, VGG-based
perceptual loss, and total variation (TV) loss. The dataset contains
8,000 images: 2,000 clean images from the Flickr Images Dataset
and 6,000 synthetic noisy images generated by applying three
noise classes—Gaussian, salt-and-pepper, and speckle—to the
clean images. The clean source images and their noisy variants
are partitioned into training, validation, and test sets using a
70/20/10 split. The experimental results indicate stable denoising
performance across all three noise types with a mean PSNR of
31.54 dB. A detailed ablation study evaluates the contribution of
each loss term. This work serves as a reproducible framework
forcombine residual learning, spectral normalization, and perceptual
losses for parameter-free restoration across the three types
of synthetic noise. Its modest computational requirements also
make it suitable for resource-constrained applications including
edge devices and medical imaging systems.