Learning from Multi-Perception Features for Real-Word Image Super-resolution

TL;DR

This paper introduces MPF-Net, a novel super-resolution method that leverages multiple perceptual features and contrastive regularization to enhance real-world image super-resolution, outperforming existing methods.

Contribution

The paper proposes a multi-perception feature extraction framework with cross-perception blocks and contrastive regularization for improved real-world image super-resolution.

Findings

Significantly outperforms state-of-the-art methods in real-world datasets.

Effectively utilizes diverse perceptual features for better reconstruction.

Contrastive regularization enhances learning capability.

Abstract

Currently, there are two popular approaches for addressing real-world image super-resolution problems: degradation-estimation-based and blind-based methods. However, degradation-estimation-based methods may be inaccurate in estimating the degradation, making them less applicable to real-world LR images. On the other hand, blind-based methods are often limited by their fixed single perception information, which hinders their ability to handle diverse perceptual characteristics. To overcome this limitation, we propose a novel SR method called MPF-Net that leverages multiple perceptual features of input images. Our method incorporates a Multi-Perception Feature Extraction (MPFE) module to extract diverse perceptual information and a series of newly-designed Cross-Perception Blocks (CPB) to combine this information for effective super-resolution reconstruction. Additionally, we introduce a contrastive regularization term (CR) that improves the model's learning capability by using newly generated HR and LR images as positive and negative samples for ground truth HR. Experimental results on challenging real-world SR datasets demonstrate that our approach significantly outperforms existing state-of-the-art methods in both qualitative and quantitative measures.