Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Syn-to-Real IQA: A Novel Perspective on Reshaping Synthetic Data Distributions

Aobo Li, Jinjian Wu, Yongxu Liu, Leida Li, Weisheng Dong

TL;DR

This work tackles the limited cross-domain generalization of BIQA models trained on synthetic distortions by identifying a discrete, clustered feature distribution that hinders regression. It introduces SynDR-IQA, a data-distribution reshaping framework with two strategies: Distribution-aware Diverse Content Upsampling (DDCUp) to boost content diversity while preserving distribution, and Density-aware Redundant Cluster Downsampling (DRCDown) to reduce overrepresented clusters and balance density. A theoretical generalization bound guides the design, highlighting the roles of diverse samples $m$ and redundancy heterogeneity $\eta$. Empirically, SynDR-IQA improves synthetic-to-authentic, synthetic-to-algorithmic, and synthetic-to-synthetic BIQA generalization across multiple backbones, with ablations confirming each component's contribution and CLIP-enhanced upsampling providing further gains. This data-centric approach offers a practical, model-agnostic route to stronger cross-domain perceptual quality assessment without incurring extra inference costs.

Abstract

Blind Image Quality Assessment (BIQA) has advanced significantly through deep learning, but the scarcity of large-scale labeled datasets remains a challenge. While synthetic data offers a promising solution, models trained on existing synthetic datasets often show limited generalization ability. In this work, we make a key observation that representations learned from synthetic datasets often exhibit a discrete and clustered pattern that hinders regression performance: features of high-quality images cluster around reference images, while those of low-quality images cluster based on distortion types. Our analysis reveals that this issue stems from the distribution of synthetic data rather than model architecture. Consequently, we introduce a novel framework SynDR-IQA, which reshapes synthetic data distribution to enhance BIQA generalization. Based on theoretical derivations of sample diversity and redundancy's impact on generalization error, SynDR-IQA employs two strategies: distribution-aware diverse content upsampling, which enhances visual diversity while preserving content distribution, and density-aware redundant cluster downsampling, which balances samples by reducing the density of densely clustered areas. Extensive experiments across three cross-dataset settings (synthetic-to-authentic, synthetic-to-algorithmic, and synthetic-to-synthetic) demonstrate the effectiveness of our method. The code is available at https://github.com/Li-aobo/SynDR-IQA.

Towards Syn-to-Real IQA: A Novel Perspective on Reshaping Synthetic Data Distributions

TL;DR

This work tackles the limited cross-domain generalization of BIQA models trained on synthetic distortions by identifying a discrete, clustered feature distribution that hinders regression. It introduces SynDR-IQA, a data-distribution reshaping framework with two strategies: Distribution-aware Diverse Content Upsampling (DDCUp) to boost content diversity while preserving distribution, and Density-aware Redundant Cluster Downsampling (DRCDown) to reduce overrepresented clusters and balance density. A theoretical generalization bound guides the design, highlighting the roles of diverse samples and redundancy heterogeneity . Empirically, SynDR-IQA improves synthetic-to-authentic, synthetic-to-algorithmic, and synthetic-to-synthetic BIQA generalization across multiple backbones, with ablations confirming each component's contribution and CLIP-enhanced upsampling providing further gains. This data-centric approach offers a practical, model-agnostic route to stronger cross-domain perceptual quality assessment without incurring extra inference costs.

Abstract

Blind Image Quality Assessment (BIQA) has advanced significantly through deep learning, but the scarcity of large-scale labeled datasets remains a challenge. While synthetic data offers a promising solution, models trained on existing synthetic datasets often show limited generalization ability. In this work, we make a key observation that representations learned from synthetic datasets often exhibit a discrete and clustered pattern that hinders regression performance: features of high-quality images cluster around reference images, while those of low-quality images cluster based on distortion types. Our analysis reveals that this issue stems from the distribution of synthetic data rather than model architecture. Consequently, we introduce a novel framework SynDR-IQA, which reshapes synthetic data distribution to enhance BIQA generalization. Based on theoretical derivations of sample diversity and redundancy's impact on generalization error, SynDR-IQA employs two strategies: distribution-aware diverse content upsampling, which enhances visual diversity while preserving content distribution, and density-aware redundant cluster downsampling, which balances samples by reducing the density of densely clustered areas. Extensive experiments across three cross-dataset settings (synthetic-to-authentic, synthetic-to-algorithmic, and synthetic-to-synthetic) demonstrate the effectiveness of our method. The code is available at https://github.com/Li-aobo/SynDR-IQA.
Paper Structure (24 sections, 1 theorem, 9 equations, 10 figures, 10 tables, 3 algorithms)

This paper contains 24 sections, 1 theorem, 9 equations, 10 figures, 10 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Problem Formalization
  5. Theoretical Analysis
  6. SynDR-IQA Framework
  7. Distribution-aware Diverse Content Upsampling
  8. Density-aware Redundant Cluster Downsampling
  9. Experiments
  10. Experimental Setups
  11. Performance Evaluation
  12. Ablation Study
  13. Visualization Analysis
  14. Conclusion
  15. UMAP Visualizations across Different Backbones and Datasets
  16. ...and 9 more sections

Key Result

Theorem 3.1

Let $\mathcal{F}$ be a hypothesis class of functions $f: \mathcal{X} \to \mathcal{Y}$, and $\mathcal{\hat{D}} = \{(x_i, y_i)\}_{i=1}^{n}$ be a dataset consisting of $m$i.i.d. samples from true distribution $\mathcal{D}$, along with their respective $k_i-1$ redundant samples drawn from the correspond where $R(f)$ is the true risk, $R_{\text{emp}}(f)$ is the empirical risk, $\operatorname{Rad}_m(\ma

Figures (10)

  • Figure 1: (a), (b), and (c) present UMAP mcinnes2018umap visualizations of the same representations learned from the synthetic distortion dataset KADID-10k kadid10k by the baseline model he2016deep. The three visualizations differ only in color mapping: (a) colors indicate the Mean Opinion Score (MOS) values (higher indicates better quality); (b) colors correspond to reference images; and (c) colors denote distortion types. Representative samples are added in (b) and (c) to illustrate the redundancy within high‑quality and low‑quality clusters, respectively. More visualizations across different backbones and datasets are provided in Appendix \ref{['MoreUmap']} to demonstrate the generality of this phenomenon.
  • Figure 2: This figure shows the core concepts of two strategies in SynDR-IQA. The DDCUp strategy (upper) selects images from candidate reference sets that are similar in distribution but diverse in content to the training reference sets for synthesizing distorted samples. The pseudo-labels of these samples depend on the nearest neighbors of their reference images. The DRCDown strategy (lower) identifies high-density clusters in the training dataset and selectively removes samples from these clusters while retaining samples from low-density regions.
  • Figure 3: UMAP visualization of features extracted from LIVEC using the same model under four different training processes: (a) trained directly on LIVEC, (b) trained directly on KADID‑10k), (c) trained on KADID‑10k based on DGQA, and (d) trained on KADID‑10k based on SynDR‑IQA.
  • Figure 4: UMAP mcinnes2018umap visualizations of features learned by VGG-16 simonyan2014very and Swin‑T liu2021swin on KADID‑10k kadid10k. Both show discrete clustering patterns on synthetic distortion data.
  • Figure 5: UMAP mcinnes2018umap visualization of ResNet-50 he2016deep trained on TID2013 2013Color, demonstrating consistent clustering trends across synthetic datasets.
  • ...and 5 more figures

Theorems & Definitions (2)

  • Theorem 3.1: Generalization Bound for Clustered Synthetic Data
  • proof