Learning a No-Reference Quality Assessment Model of Enhanced Images With Big Data

Ke Gu, Dacheng Tao, Junfei Qiao, Weisi Lin

Introduction

Traditional Image Quality Assessment (IQA) metrics are typically designed to detect degradations such as noise, blur, or compression artifacts. In these contexts, the original captured image is treated as the ground truth. However, image enhancement—encompassing techniques like dehazing, low-light correction, and contrast boosting—aims to produce an image that is perceptually superior to the original. Standard metrics often penalize these improvements because they deviate from the original's statistics. Furthermore, most No-Reference (NR) metrics suffer from poor generalization due to training on limited human-labeled datasets.

To address these challenges, this research introduces BIQME (Blind Image Quality Measure of Enhancement), a robust, opinion-unaware metric trained on a massive dataset. The authors further demonstrate its utility through BOIEM, a quality-guided enhancement framework that optimizes parameters to maximize perceived quality.

Visual comparison of original (top) and enhanced (bottom) images across different categories like low-light, hazy, and contrast-distorted.

The BIQME Model

The BIQME model is built upon a sophisticated feature extraction process and a novel "big data" training strategy. It extracts 17 hand-crafted features across five critical perceptual dimensions:

  1. Contrast: Evaluated using phase congruence and entropy.
  2. Sharpness: Measured via wavelet subband log-energy.
  3. Brightness: Assessed through the entropy of luminance-varying versions of the image.
  4. Colorfulness: Derived from saturation and global mean/variance.
  5. Naturalness: Captured using Natural Scene Statistics (NSS) and the Dark Channel Prior (DCP).

To overcome the bottleneck of human subjective testing (Mean Opinion Scores), the authors employed a "model distillation" strategy. They used a high-performance Full-Reference metric (C-PCQI) to generate pseudo-labels for over 100,000 simulated enhanced images. This allowed for the training of a Support Vector Regression (SVR) module on a dataset orders of magnitude larger than previous efforts, ensuring high generalization and robustness.

The BOIEM Framework

Beyond just measuring quality, the authors developed the BIQME-Optimized Image Enhancement Method (BOIEM). This framework treats image enhancement as an optimization problem. It utilizes a cascade of histogram modification techniques, such as modified Adaptive Gamma Correction with Weighting Distribution (AGCWD).

Instead of relying on fixed parameters, BOIEM iteratively searches for the specific brightness and contrast settings that yield the highest BIQME score for a given input. By incorporating naturalness features like the Dark Channel Prior, the system effectively penalizes over-enhancement, preventing the "cartoonish" or over-saturated artifacts common in automated tools.

Experimental Evaluation

The performance of BIQME was rigorously tested across nine diverse datasets, including contrast-specific databases (CID2013, CCID2014) and general IQA subsets (TID2008, CSIQ, TID2013, SIQAD).

As shown in the comparative results, BIQME consistently outperformed state-of-the-art NR-IQA metrics like NIQE and BRISQUE. In many instances, its accuracy approached that of Reduced-Reference metrics, despite having no access to the original image. Furthermore, the model is highly efficient, processing a standard 768x576 image in under one second.

Conclusion

By leveraging big data and a multi-dimensional feature set, BIQME provides a reliable way to evaluate the quality of enhanced images without a reference. The accompanying BOIEM framework demonstrates how such a metric can be used to automate high-quality image enhancement. Future work may explore replacing hand-crafted features with deep learning architectures to further improve performance, provided the large-scale training approach is maintained to prevent overfitting.

Summary updated

BibTeX
@article{gu2018learning,
  author = {Ke Gu and Dacheng Tao and Junfei Qiao and Weisi Lin},
  title = {{Learning a No-Reference Quality Assessment Model of Enhanced Images With Big Data}},
  journal = {IEEE Trans. Neural Networks Learn. Syst.},
  year = {2018},
  doi = {10.1109/tnnls.2017.2649101},
}
Figure