Curiously Effective Features for Image Quality Prediction

Sören Becker, Thomas Wiegand, Sebastian BosseICIP 2021 · DOI

Introduction

In the field of Image Quality Assessment (IQA), the prevailing paradigm has long been that predictive models must rely on "perceptually relevant" features. These features are typically either handcrafted by experts to simulate the human visual system (HVS) or learned from vast datasets of natural images to capture natural scene statistics (NSS). This paper challenges this fundamental assumption by demonstrating that features derived from pure random noise can predict image quality nearly as effectively as sophisticated, learned models, provided the model operates in a highly overparameterized regime.

Methodology

The researchers conducted their investigation using the CORNIA (Codebook Representation for No-Reference Image Assessment) framework. Traditionally, CORNIA extracts small 7x7 patches from images and learns a "codebook" of representative visual patterns via k-means clustering. An image is then encoded based on this codebook, and a linear Support Vector Regression (SVR) is used to predict the Mean Opinion Score (MOS).

In this study, the authors replaced the learned codebook with several stochastic baselines:

  • Random Patches: Unprocessed patches sampled directly from natural images.
  • Pure Noise: Codebooks generated from independent and identically distributed (i.i.d.) samples of Normal, Laplace, or Uniform distributions. These noise-based features contain zero prior information about human vision or the structure of natural scenes.

The models were evaluated on major IQA benchmarks, including the LIVE, TID2013, and CSIQ datasets, to test both internal performance and cross-database generalization.

The Power of Overparameterization

A central contribution of this work is the observation of the Double Descent phenomenon in the context of IQA. As the number of features increases, the model's performance (measured by Pearson Correlation Coefficient, or PCC) initially improves. However, performance degrades as the number of features approaches the number of training samples—the "interpolation point"—before improving again as the model becomes "highly overparameterized."

Figure 1: PCC performance vs. number of features. The plot illustrates the double descent phenomenon and shows that noise-based models (Normal, Laplace) eventually match the performance of the learned CORNIA model as the number of features exceeds 2,000.

The "noise" models only become "curiously effective" in this overparameterized regime. This suggests that while a single learned feature is significantly more informative than a single noise feature, a massive ensemble of random features can collectively capture the necessary signal to match the performance of specialized models.

Spectral Convergence and the "Lottery Ticket"

To explain why random noise can be so effective, the authors propose a "lottery effect." When generating thousands of random feature extractors, a subset will, by pure chance, align with the band-pass characteristics of the human visual system.

The paper uses spectral analysis to support this. When the SVR selects the most "important" features (those with the highest absolute coefficients), the power spectra of these noise-derived filters actually begin to resemble the human contrast sensitivity function. The optimization process "finds" the relevant filters even within the chaos of random noise.

Figure 3: Normalized power spectra of features ranked by SVR importance. The plots show how different feature sets (CORNIA, patches, and various noise distributions) distribute power across spatial frequencies.

Key Results and Discussion

On the LIVE database, the standard CORNIA model achieved a PCC of 0.93. Remarkably, models using Normal and Laplace noise reached the same level of performance once the feature count exceeded approximately 2,000. Furthermore, in cross-database tests (e.g., training on LIVE and testing on TID2013), the noise-based models demonstrated generalization capabilities on par with learned models. This indicates that the effectiveness of random features is a robust property rather than an artifact of a specific dataset.

Conclusion and Future Work

This research suggests that the success of modern IQA models may be driven more by the quantity of features in a high-dimensional space than by the specific quality or "perceptual relevance" of those features. While using 10,000 noise features is computationally intensive, these findings provide a new perspective on model complexity and the minimal requirements for image quality prediction.

Future research should investigate whether these findings extend to deep convolutional neural networks (CNNs) and if random features are equally effective in Full-Reference (FR) IQA scenarios, where a reference image is available for comparison.

Summary updated

BibTeX
@inproceedings{becker2021curiously,
  author = {Sören Becker and Thomas Wiegand and Sebastian Bosse},
  title = {{Curiously Effective Features for Image Quality Prediction}},
  booktitle = {ICIP},
  year = {2021},
  doi = {10.1109/icip42928.2021.9506756},
}
Figure