A New Dataset and Framework for Real-World Blurred Images Super-Resolution

TL;DR

The paper tackles the prevalence of intentional blur in real-world images and its detrimental effect on blind SR. It introduces ReBlurSR, a blur-focused dataset with 2,931HQ images (plus synthetic samples) and blur-region annotations, and proposes PBaSR, a blur-adaptive framework built from two modules: CDM to disentangle blur and general data, and CFM to fuse priors via model interpolation without increasing inference cost. Extensive cross-dataset and ablation experiments show PBaSR delivers state-of-the-art perceptual-quality improvements on blur images (LPIPS improvements of roughly $0.02$–$0.10$) while preserving performance on general SR tasks, and it generalizes across multiple SR architectures. The work provides a practical path to robust blur-aware SR and offers publicly available code to reproduce results.

Abstract

Recent Blind Image Super-Resolution (BSR) methods have shown proficiency in general images. However, we find that the efficacy of recent methods obviously diminishes when employed on image data with blur, while image data with intentional blur constitute a substantial proportion of general data. To further investigate and address this issue, we developed a new super-resolution dataset specifically tailored for blur images, named the Real-world Blur-kept Super-Resolution (ReBlurSR) dataset, which consists of nearly 3000 defocus and motion blur image samples with diverse blur sizes and varying blur intensities. Furthermore, we propose a new BSR framework for blur images called Perceptual-Blur-adaptive Super-Resolution (PBaSR), which comprises two main modules: the Cross Disentanglement Module (CDM) and the Cross Fusion Module (CFM). The CDM utilizes a dual-branch parallelism to isolate conflicting blur and general data during optimization. The CFM fuses the well-optimized prior from these distinct domains cost-effectively and efficiently based on model interpolation. By integrating these two modules, PBaSR achieves commendable performance on both general and blur data without any additional inference and deployment cost and is generalizable across multiple model architectures. Rich experiments show that PBaSR achieves state-of-the-art performance across various metrics without incurring extra inference costs. Within the widely adopted LPIPS metrics, PBaSR achieves an improvement range of approximately 0.02-0.10 with diverse anchor methods and blur types, across both the ReBlurSR and multiple common general BSR benchmarks. Code here: https://github.com/Imalne/PBaSR.

Figures (23)

• Figure 1: Recent BSR methods' performance on blur data. Left: Examples of real images with blur. Right: LPIPS on blur and non-blur images in DIV2K-Val.
• Figure 2: Data collection and partition of ReBlurSR. (a) Three methods of collecting the ReBlurSR dataset. (b) The samples from different data partitions.
• Figure 3: Data analysis of the ReBlurSR dataset. (a) The sample distribution of image resolution and blur type. (b) The local gradient in the blur region of different blur intensity subsets. (c) The sample distribution and the HR images' quality assessment based on NIQE niqe of different blur area sizes and blur intensities in ReBlurSR.
• Figure 4: (a) The performance of FeMaSR femasr_chen2022real fine-tuned on different proportions of general (G) and blur (B) data (keeping the total training data amount the same). (b) FeMaSR's performance during late iterations which were unified trained on the combination of all general and blur images (darker colors denote later iterations). (c) The adversarial loss of general discriminator on the HR images of ReblurSR-Test.
• Figure 5: The structure of the Cross Disentanglement Module (CDM).