No-reference Image Quality Assessment With A Gradient-induced Dictionary

Leida Li, Dong Wu, Jinjian Wu, Jiansheng Qian, Beijing ChenKSII Trans. Internet Inf. Syst 2016 · DOI · OpenReview

Abstract

Evaluating image quality without a reference (NR-IQA) remains a challenge due to the complexity of human visual perception. This paper presents GRID (GRadient-Induced Dictionary), a novel NR-IQA metric that leverages image gradients for dictionary learning. By focusing on structural information rather than raw pixel intensities, GRID achieves high prediction accuracy with a compact dictionary of only 800 atoms. Experimental results on benchmark datasets demonstrate that GRID outperforms existing methods in both accuracy and computational efficiency, offering a robust solution for real-time quality monitoring.

Introduction

No-Reference Image Quality Assessment (NR-IQA) is critical for applications where a pristine reference image is unavailable. Traditional methods often rely on Natural Scene Statistics (NSS) or large visual codebooks. However, these approaches face two primary hurdles: structural sensitivity and computational efficiency. Distortions like blur or noise primarily degrade image structures, which raw pixel patches may fail to capture efficiently. Furthermore, high-performing dictionary-based methods often require massive codebooks (e.g., 10,000 dimensions), making them impractical for resource-constrained applications.

The GRID Framework

The authors propose a two-stage framework that utilizes a gradient-domain dictionary to capture structural degradations. The process involves feature extraction via a learned dictionary followed by a two-stage prediction model.

Overview of the GRID framework for model training and quality assessment.

Gradient-Induced Dictionary Learning

Unlike previous methods that learn from raw pixels, GRID computes image gradients using the Scharr operator. Patches are extracted from natural images, normalized, and clustered using K-means to form a dictionary. This approach allows a very small dictionary of just 800 atoms to represent common structural patterns like edges and ridges effectively.

Natural images used for training and the resulting gradient-induced dictionary atoms.

Feature Extraction and Prediction

For a given test image, features are extracted by calculating the Euclidean distance between image patches and the dictionary atoms (Euclidean-norm coding). Max-pooling is then applied to generate a final feature vector. The prediction stage uses a Support Vector Machine (SVM) to classify distortion types and Support Vector Regression (SVR) to predict quality scores. The final score is a weighted sum of these regression outputs based on distortion probabilities.

Experimental Results

GRID was rigorously evaluated on benchmark databases including LIVE, CSIQ, TID2008, and TID2013.

  • Accuracy: On the LIVE database, GRID achieved a Pearson Linear Correlation Coefficient (PLCC) of 0.9735 and a Spearman Rank Order Correlation Coefficient (SRCC) of 0.9722, surpassing established metrics like BRISQUE and CORNIA.
  • Generalization: In cross-database tests (training on LIVE, testing on TID2013), GRID maintained a PLCC of 0.9131, indicating robust generalization capabilities.
  • Efficiency: GRID is approximately 9 times faster than CORNIA. The use of the gradient domain allows performance to saturate at just 800 atoms, whereas raw-pixel methods require up to 10,000 atoms for similar accuracy.

Conclusion

GRID provides a highly efficient and accurate solution for NR-IQA by prioritizing structural information through gradient-domain dictionary learning. While the current model excels at structural distortions, its performance on contrast-distorted images remains a limitation. Future work will focus on incorporating color information and improving robustness for multiply-distorted images. The compact nature of the GRID dictionary makes it a strong candidate for real-time image quality assessment.

Summary updated

BibTeX
@article{li2016no,
  author = {Leida Li and Dong Wu and Jinjian Wu and Jiansheng Qian and Beijing Chen},
  title = {{No-reference Image Quality Assessment With A Gradient-induced Dictionary}},
  journal = {KSII Trans. Internet Inf. Syst.},
  year = {2016},
  doi = {10.3837/tiis.2016.01.017},
}
Figure