Fortifying Fully Convolutional Generative Adversarial Networks for Image Super-Resolution - Using Divergence Measures

Aug 23, 2026

Super-Resolution (SR) aims to transform a Low-Resolution (LR) image into a High-Resolution (HR) counterpart, despite the LR-to-HR mapping being non-invertible and lossy. This work introduces SuRGe, a fully-convolutional GAN architecture for SR that combines convolutional features from multiple generator depths using learnable convex weights. It employs Jensen–Shannon and GromovWasserstein losses on SR–HR and LR–SR distribution pairs, and trains the discriminator with Wasserstein loss plus gradient penalty to mitigate mode collapse. SuRGe delivers improved end-to-end performance with low inference time, outperforming 28 state-of-the-art methods on 10 benchmark datasets.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages20

Reading Time50 min

AccessFree to download

Introduction

Motivation and limitations of classical methods

GAN-based super-resolution background

Proposed SuRGe generator design and feature mixing

Divergence-based losses and discriminator training

1 arXiv :2404 .06294v2 [ ee ss .IV] 3 Jul 2026 Fortifying Fully Convolutional Generative Adversarial Networks for Image Super-Resolution Using Divergence Measures Arkaprabha Basu, Kushal Bose, Sankha Subhra Mullick, Anish Chakrabarty, and Swagatam Das Abstract Super-Resolution (SR) is a time-hallowed image processing problem that aims to improve the quality of a Low-Resolution (LR) sample up to the standard of its High-Resolution (HR) counterpart. We aim to address this by introducing Super-Resolution Generator (SuRGe), a fully-convolutional Generative Adversarial Network (GAN)-based architecture for SR. We show that distinct convolutional features obtained at increasing depths of a GAN generator can be optimally combined by a set of learnable convex weights to improve the quality of generated SR samples. In the process, we employ the Jensen–Shannon and the GromovWasserstein losses respectively between the SR-HR and LR-SR pairs of distributions to further aid the generator of SuRGe to better exploit the available information in an attempt to improve SR. Moreover, we train the discriminator of SuRGe with the Wasserstein loss with gradient penalty, to primarily prevent mode collapse. The proposed SuRGe, as an end-to-end GAN workflow tailor-made for super-resolution, offers improved performance while maintaining low inference time. The efficacy of SuRGe is substantiated by its superior performance compared to 28 state-of-the-art contenders on 10 benchmark datasets. Index Terms Generative Adversarial Networks, Image Super-Resolution, Convolutional Neural Networks, Divergence Measures I. INTRODUCTION A Low-resolution (LR) image sacrifices information of its high-resolution (HR) counterpart in favor of general utility, such as displaying or editing on smaller screens, low storage requirements, and fast transmission. Super-resolution attempts to recover the original HR copy from an LR input. However, the initial HR to LR transformation is commonly non-invertible and lossy [1] . Thus, recovering the HR by estimating a Super Resolution (SR) analog is an ill-posed problem with the risk of a distorted output [2] . The classical interpolation methods for super-resolution only exploit local information and are thus incapable of generating commendable SR [3] . While global image features extracted by the deep convolutional networks translated to a much-improved performance [4], [5], limited generalizability and distorted, SR remains a major concern [6] . The landscape of super-resolution techniques had a significant breakthrough with the advent of the Generative Adversarial Network (GAN) [7] . A super-resolution GAN [8] embraces the canonical two-player adversarial game between a generator G and a discriminator D with some minor modifications. Specifically, in a super-resolution task, G attempts to map an LR input to an HR ground truth, generating an estimated SR in the process. The discriminator D helps G by providing adversarial feedback by distinguishing between an HR ground truth and G-generated SR. While GAN-based super-resolution offers generalizability through their generative power, they often have SR outputs that lose finer details or are plagued by artifacts [9], [10] . In this paper, we propose a GAN-based super-resolution method called Super-Resolution Generator (SuRGe) 1. In a superresolution task, to generate a good quality SR image, it is necessary to consider both the low-level local features (for example, colors, textures, edges, etc.) and the high-level global ones (such as individual object shapes, relative positioning of objects and background, object orientation, etc.) . As noted in [14], [15], higher-level global features are progressively captured by convolutional filters residing deeper in the network. Taking inspiration from [14],[16] in the generator G of the proposed SuRGe, we preserve the hierarchically complex features and dictate their flow through skip connections. However, skip conne

FAQs

What problem does the paper address in image processing?

The paper addresses image super-resolution: recovering a high-resolution image from a low-resolution input while avoiding distorted outputs caused by the lossy, non-invertible LR-to-HR transformation.

How does SuRGe combine features to improve super-resolution quality?

SuRGe uses learnable convex weights to optimally combine convolutional features extracted at increasing depths of the GAN generator, guided by a mixing-module design and selective use of skip connections.

Which loss functions does SuRGe use and why?

It uses Jensen–Shannon and GromovWasserstein losses between distribution pairs to better exploit available information, and trains the discriminator with Wasserstein loss with gradient penalty to primarily prevent mode collapse.

Aug 23, 2026

Upload by Lucas Martin