Blind Image Quality Assessment With Joint Entropy Degradation

Weiping Ji, Jinjian Wu, Man Zhang, Zuozhi Liu, Guangming Shi, Xuemei XieIEEE Access 2019 · OpenReview

Abstract

Blind image quality assessment (BIQA) aims to evaluate the quality of an image without a pristine reference, a task essential for perception-oriented image processing. This paper proposes assessing image quality through visual entropy degradation, grounded in information theory. By extracting neuroscience-inspired features—local intensity, gradient, and orientation—and calculating their joint statistical distributions, the proposed method quantifies the loss of visual content. Experimental results on the LIVE, CSIQ, and TID2013 databases demonstrate that the proposed method achieves superior performance compared to state-of-the-art BIQA methods.

Introduction: The Information Theory of Quality

In digital imaging, evaluating quality without a reference image (Blind Image Quality Assessment or BIQA) is a significant challenge. Traditional methods often struggle with the diverse ways different distortions—such as blur, noise, or compression—affect visual content. This research approaches the problem through the lens of information theory, positing that distortions essentially degrade the "visual entropy" of an image.

A pristine image with high visual information content.
A distorted version of the same image, exhibiting a loss of visual entropy through compression artifacts.

As illustrated in the comparison above, distortions introduce artifacts that alter the statistical regularity of natural scenes. By measuring the degradation of joint statistical distributions of key visual features, the proposed model can objectively quantify quality in a manner that aligns closely with human subjective perception.

Methodology: Capturing Visual Content

The proposed framework utilizes a three-step process to map statistical degradation to quality scores:

  1. Neuroscience-Inspired Feature Extraction: Drawing from the function of "simple cells" in the primary visual cortex, the model extracts three local features:
  2. Local Intensity: Normalized luminance to remove redundant information.
  3. Local Gradient: Captured via Gaussian difference filters to represent structural changes.
  4. Local Orientation Selectivity: Modeled using Orientation Selectivity based Visual Patterns (OSVP) to represent edges and textures.
  5. Joint Statistical Distribution: Rather than analyzing features in isolation, the method calculates their joint probability distributions. This captures the high correlation between features in natural images, making the measure more sensitive to subtle distortions than univariate analysis.
  6. Quality Mapping: A Support Vector Regression (SVR) module is employed to learn the mapping between these complex statistical distributions and human-assigned Mean Opinion Scores (MOS).

Experimental Results and Performance

The model was rigorously evaluated on three major benchmark databases: LIVE, CSIQ, and the extensive TID2013 dataset, which contains 3,000 images across 24 distortion types.

  • State-of-the-Art Accuracy: On the TID2013 database, the proposed method achieved a Spearman Rank Correlation Coefficient (SRCC) of 0.960, significantly outperforming established models like BRISQUE (0.893) and NIQE (0.845).
  • Robustness: In cross-database validation (e.g., training on LIVE and testing on CSIQ), the model maintained high accuracy with an SRCC of 0.916, proving it generalizes well to unseen images and diverse noise types.
  • Consistency: The approach proved particularly effective for common distortions like JPEG compression and Gaussian blur, matching or exceeding the performance of many "Full-Reference" models that require access to the original image.

Conclusion and Future Directions

By linking image quality to the degradation of visual entropy, this research provides a robust, biologically-inspired framework for blind quality assessment. While the current implementation faces computational challenges due to the high dimensionality of joint distributions—requiring feature discretization into bins—it sets a new benchmark for BIQA performance. Future work may involve integrating the visual entropy concept into deep learning architectures to further automate feature extraction and improve processing efficiency.

Summary updated

Abstract

Blind image quality assessment (BIQA) aims to evaluate the quality of an image without pristine image objectively, which is highly desired in many perception-oriented image processing systems. Distortions degrade the visual contents and cause image quality degradation. Moreover, the visual contents of an image suffer from individual degradations by different types and different levels of distortions, which makes us difficult to analyze the quality degradation. From the perspective of information theory, there is a decrease in the amount of the visual contents when images are distorted. Therefore, in this paper, the image quality is assessed through its visual entropy degradation. Researches on the neuroscience indicate that the simple cells in the local receptive field can be characterized as being spatially localized and oriented, then the local intensity, gradient, and orientation features are extracted to represent the visual contents of an image. By deducing the joint entropy equation, the joint entropy is related to the statistical distributions. Next, in order to measure the visual entropy, the joint statistical distributions of those features are calculated. Finally, by measuring the degradations on these distributions of distorted images, a novel BIQA method is proposed. The experimental results on the databases of LIVE, CSIQ, and TID2013 show that the proposed method has superior performance than other state-of-the-art BIQA methods.
BibTeX
@inproceedings{ji2019blind,
  author = {Weiping Ji and Jinjian Wu and Man Zhang and Zuozhi Liu and Guangming Shi and Xuemei Xie},
  title = {{Blind Image Quality Assessment With Joint Entropy Degradation}},
  booktitle = {IEEE Access},
  year = {2019},
  url = {https://openreview.net/forum?id=-XBl7hCgJz},
}
Figure