SRTGAN: Triplet Loss based Generative Adversarial Network for Real-World Super-Resolution

TL;DR

This paper introduces SRTGAN, a novel triplet loss-based GAN for real-world super-resolution that leverages LR images as negative samples, improving the quality of high-resolution images in practical applications.

Contribution

The paper proposes a new triplet-based adversarial loss function and a fusion of multiple losses to enhance super-resolution quality on real-world degraded images.

Findings

Outperforms existing methods on the RealSR dataset

Achieves higher perceptual fidelity in super-resolved images

Demonstrates improved differentiation between HR and LR images

Abstract

Many applications such as forensics, surveillance, satellite imaging, medical imaging, etc., demand High-Resolution (HR) images. However, obtaining an HR image is not always possible due to the limitations of optical sensors and their costs. An alternative solution called Single Image Super-Resolution (SISR) is a software-driven approach that aims to take a Low-Resolution (LR) image and obtain the HR image. Most supervised SISR solutions use ground truth HR image as a target and do not include the information provided in the LR image, which could be valuable. In this work, we introduce Triplet Loss-based Generative Adversarial Network hereafter referred as SRTGAN for Image Super-Resolution problem on real-world degradation. We introduce a new triplet-based adversarial loss function that exploits the information provided in the LR image by using it as a negative sample. Allowing the patch-based discriminator with access to both HR and LR images optimizes to better differentiate between HR and LR images; hence, improving the adversary. Further, we propose to fuse the adversarial loss, content loss, perceptual loss, and quality loss to obtain Super-Resolution (SR) image with high perceptual fidelity. We validate the superior performance of the proposed method over the other existing methods on the RealSR dataset in terms of quantitative and qualitative metrics.