Less is More: Learning Reference Knowledge Using No-Reference Image Quality Assessment

Xudong Li, Yan Zhang, Shengchuan Zhang, Sicheng Zhao, Ke Li, Liujuan Cao, Rongrong Ji

Keywords: Image Quality Assessment, Inductive Bias Regularization, Reference Knowledge

Abstract: Image Quality Assessment (IQA) with reference images has achieved great success by imitating the human vision system, in which the image quality is effectively assessed by comparing the query image with its pristine reference image. However, for the images in the wild, it is quite difficult to access accurate reference images. We argue that it is possible to learn reference knowledge under the \emph{No-Reference Image Quality Assessment} (NR-IQA) setting, which is effective and efficient empirically. Concretely, by innovatively introducing a novel feature distillation method in IQA, we propose a new framework to learn comparative knowledge from non-aligned reference images. Then, we further propose inductive bias regularization to inject different inductive biases into the model to achieve fast convergence and avoid overfitting. Such a framework not only solves the congenital defects of NR-IQA but also improves the feature extraction framework, enabling it to express more abundant quality information. Surprisingly, our method utilizes less input—eliminating the need for reference images during inference—while obtaining more performance compared to some IQA methods that do require reference images. Comprehensive experiments on eight standard IQA datasets show that our approach outperforms state-of-the-art NR-IQA methods.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2007