HDIQA: A Hyper Debiasing Framework for Full Reference Image Quality Assessment

Mingliang ZHOU*, Heqiang WANG, Xuekai WEI, Yong FENG, Jun LUO, Huayan PU, Jinglei ZHAO, Liming WANG, Zhigang CHU, Xin WANG, Bin FANG, Zhaowei SHANG

*Corresponding author for this work

Research output: Journal PublicationsJournal Article (refereed)peer-review


Recent methods that project images into deep feature spaces to evaluate quality degradation have produced inefficient results due to biased mappings; i.e., these projections are not aligned with the perceptions of humans. In this paper, we develop a hyperdebiasing framework to address such bias in full-reference image quality assessment. First, we perform orthogonal Tucker decomposition on the top of feature tensors extracted by a feature extraction network to project features into a robust content-agnostic space and effectively eliminate the bias caused by subtle image perturbations. Second, we propose a hypernetwork in which the content-aware parameters are produced for reprojecting features in a deep subspace for quality prediction. By leveraging the content diversity of large-scale blind-reference datasets, the perception rule between image content and image quality is established. Third, a quality prediction network is proposed by combining debiased content-aware and content-agnostic features to predict the final image quality score. To demonstrate the efficacy of our proposed method, we conducted numerous experiments on comprehensive databases. The experimental results validate that our method achieves state-of-the-art performance in predicting image quality.

Original languageEnglish
Pages (from-to)545-554
Number of pages10
JournalIEEE Transactions on Broadcasting
Issue number2
Publication statusPublished - Jun 2024
Externally publishedYes

Bibliographical note

Publisher Copyright:
© 1963-12012 IEEE.


  • deep feature
  • hypernetwork
  • Image quality assessment
  • Tucker decomposition


Dive into the research topics of 'HDIQA: A Hyper Debiasing Framework for Full Reference Image Quality Assessment'. Together they form a unique fingerprint.

Cite this