Source-linked AI summary
No-Reference Color Image Quality Assessment: From Entropy to Perceptual Quality
Xiaoqiao Chen, Qingyi Zhang, Manhui Lin, Guangyi Yang, Chu He
TL;DR
Image quality assessment often must operate without a reference image, while existing general-purpose methods may underrepresent color and spatial feature distributions. ENIQA combines entropy and mutual-information features with a two-stage prediction framework, and reports superior overall performance on LIVE and TID2013.
Problem
General-purpose NR-IQA methods may underconsider color information or ignore spatial distributions, while reference images are often unavailable in practice.
Source: §1 · pp. 2–3Method
ENIQA extracts two-dimensional entropy and mutual-information features in spatial and frequency domains, then uses SVC for distortion classification and SVR for quality prediction.
Source: §1 · p. 3, §4 · p. 10Results
ENIQA showed superior overall performance on LIVE and TID2013 in cross-validation and cross-database experiments.
Source: §4 · p. 10Takeaways & Limitations
ENIQA assesses different image distortion categories and shows high consistency between objective and subjective assessment.
Source: §1 · p. 3
Abstract
from arXiv · showhide
This paper presents a high-performance general-purpose no-reference (NR) image quality assessment (IQA) method based on image entropy. The image features are extracted from two domains. In the spatial domain, the mutual information between the color channels and the two-dimensional entropy are calculated. In the frequency domain, the two-dimensional entropy and the mutual information of the filtered sub-band images are computed as the feature set of the input color image. Then, with all the extracted features, the support vector classifier (SVC) for distortion classification and support vector regression (SVR) are utilized for the quality prediction, to obtain the final quality assessment score. The proposed method, which we call entropy-based no-reference image quality assessment (ENIQA), can assess the quality of different categories of distorted images, and has a low complexity. The proposed ENIQA method was assessed on the LIVE and TID2013 databases and showed a superior performance. The experimental results confirmed that the proposed ENIQA method has a high consistency of objective and subjective assessment on color images, which indicates the good overall performance and generalization ability of ENIQA. The source code is available on github https://github.com/jacob6/ENIQA.
1 Introduction
Image distortion is unavoidable across acquisition and transmission, and image quality affects people’s subjective experience and information acquisition. Because reference images are often unavailable, the introduction motivates general-purpose no-reference assessment and identifies gaps in color and spatial-feature handling that ENIQA addresses.
1 Introduction
- Reference images are often inaccessible in practice, motivating no-reference methods that estimate quality from features of the distorted image.The introduction describes NR-IQA as having practical value where a reference image cannot be supplied.Source: §1 · p. 2
- Application-specific NR-IQA methods require prior knowledge of distortion type, while deep-learning methods face limited IQA training data and less interpretable features.The introduction presents manually extracted features as an effective and accurate way to summarize image distortion.Source: §1 · p. 2
- Existing general-purpose NR-IQA methods may underuse color information and pixel-feature spatial distributions, motivating richer image representations.The introduction notes that some entropy-based methods aggregate grayscale statistics without capturing spatial feature distributions.Source: §1 · p. 3
- ENIQA combines spatial and frequency-domain two-dimensional entropy and mutual information features, with saliency-weighted patches and log-Gabor filtering, then uses SVC and SVR for quality prediction.The method is evaluated on LIVE and TID2013, with the introduction reporting high consistency between objective and subjective assessment.Source: §1 · p. 3
2 The Proposed ENIQA Framework
ENIQA extracts entropy and mutual-information features from color images in spatial and frequency domains, then combines distortion classification with distortion-specific regression to predict quality. Its feature design uses two-dimensional entropy to represent local grayscale structure, channel relationships, and log-Gabor sub-bands.
2.2 Mutual Information
- ENIQA computes mutual information between the R, G, and B channels to represent their statistical relationships.The paper also computes MI between sub-band images across orientations and between the two center frequencies.Source: §2 · p. 3, §2.2 · p. 5
2 The Proposed ENIQA Framework
- ENIQA combines 56 features from five groups with SVC distortion probabilities and distortion-specific regression scores to produce a weighted quality score.The feature groups cover color-channel MI, spatial TE statistics, sub-band TE statistics, and sub-band MI, extracted at two scales.Source: §2 · p. 3
2.1 Two-Dimensional Entropy
- TE represents spatial correlation by pairing a pixel's grayscale value with the mean grayscale value of its neighbors.The implementation uses the eight adjacent neighbors and summarizes local TE values with their mean and skewness, using the 80% more important patches.Source: §2.1 · pp. 3–5
- TE distributions distinguish distortion type and level in the illustrated LIVE examples, supporting TE as a quality feature.The figure compares histograms across distortion types at similar DMOS and across Gaussian-blur distortion levels.Source: §2.1 · pp. 4–5, Figure 3 · p. 5
2.3 Log-Gabor Filtering
- A log-Gabor filter bank with four orientations and two frequency bands produces eight sub-band images for frequency-domain features.The filter's transfer function is parameterized by center frequency, center orientation, and frequency and orientation widths.Source: §2.3 · p. 6
3 Experimental Results
Experiments assess ENIQA on LIVE and test cross-database generalization on shared distortion types in TID2013. The results examine feature discrimination, benchmark performance, window size, classification, and statistical comparisons.
3 Experimental Results
- Evaluation uses LIVE and TID2013; cross-database testing is limited to their four shared distortion categories, and performance is measured with SROCC, PLCC, and RMSE.LIVE contains 779 distorted images across five categories; testing on TID2013 uses its 24 natural reference images.Source: §3 · p. 6
3.1 Correlation of Feature Vectors with Human Opinion
- Feature-combination plots distinguish JPEG and WN well, separate GBlur to some extent, but show overlap among lower-distortion GBlur, FF, and JP2K.The plotted axes are feature combinations; points represent LIVE images, with markings for distortion type and colors for image scores.Source: §3.1 · p. 6, Figure 4 · p. 7
3.2 Correlation of Individual Feature Vectors with Human Perception
- Among individual feature groups, two-dimensional entropy contributes most, sub-band mutual information follows, and color-channel mutual information contributes least but extends entropy features.The groups are evaluated using 1,000 train-test trials, with median SROCC, PLCC, and RMSE reported.Source: §3.2 · p. 7, Table 2 · p. 7, Table 3 · p. 7
3.3 Comparison with Other IQA Methods
- ENIQA ranks second overall among NR methods on LIVE, leads SROCC for JPEG and WN, and outperforms some classic FR methods such as SSIM.For JPEG and WN, its PLCC and RMSE are better than those of all other NR methods except BRISQUE.Source: §3.3 · pp. 7–8, Table 5 · p. 8, Table 6 · p. 8
3.4 Variation with Window Size
- SROCC peaks at an 8 × 8 window, while runtime mostly decreases as window size grows; at 32 × 32, overall SROCC remains above 0.9.The authors select K = L = 8 as a compromise, while noting larger windows can trade accuracy for real-time performance.Source: §3.4 · p. 8, Figure 5 · p. 8, Table 8 · p. 8
3.5 Statistical Significance Testing
- On median SROCC, ENIQA exceeds all compared methods except BRISQUE, although BRISQUE and SSEQ are statistically superior to ENIQA in the t-tests.The comparisons use 1,000 trials; the t-test null hypothesis compares mean correlations at the 95% confidence level.Source: §3.5 · pp. 8–9, Figure 6 · p. 9, Table 9 · p. 10
3.6 Classification Performance Analysis
- With 56 features, JP2K classification accuracy reaches 71.6369%; WN is rarely confused with other categories, while the other four distortions are more confusable.The similar TE distributions of JP2K, GBlur, and FF correspond to their greater confusion.Source: §3.6 · p. 9, Figure 7 · p. 9, Table 10 · p. 10
3.7 Database Independence
- Trained on LIVE and tested on four shared distortion types in TID2013, ENIQA outperforms every compared NR method, including BRISQUE.The LIVE and TID2013 prediction scatter points show a strong linear relationship with DMOS/MOS under the reported testing protocols.Source: §3.7 · p. 9, Table 11 · p. 10, Figure 8 · p. 10
4 Conclusions
ENIQA is a general-purpose NR-IQA method that combines entropy-derived color-image features with SVC/SVR prediction. Across evaluations on LIVE and TID2013, it showed superior overall performance and generalization ability relative to other state-of-the-art methods.
4 Conclusions
- ENIQA extracts mutual information and two-dimensional entropy features in both spatial and frequency domains, combining log-Gabor filtering and saliency detection.SVC and SVR are used to predict the image quality score.Source: §4 · p. 10
- Cross-validation and cross-database experiments compared ENIQA with several full- and no-reference IQA approaches on LIVE and TID2013.Source: §4 · p. 10
- ENIQA showed superior overall performance and generalization ability compared with other state-of-the-art methods.Source: §4 · p. 10