Abstract
Blind image quality assessment (BIQA) aims to predict perceptual image quality scores without access to reference images. State-of-the-art BIQA methods typically require subjects to score a large number of images to train a robust model. However, subjective quality scores are imprecise, biased, and inconsistent, and it is challenging to obtain a large scale database, or to extend existing databases, because of the inconvenience of collecting images, training the subjects, conducting subjective experiments, and realigning human quality evaluations. To combat these limitations, this paper explores and exploits preference image pairs (PIPs) such as "the quality of image $I_a$ is better than that of image $I_b$" for training a robust BIQA model. The preference label, representing the relative quality of two images, is generally precise and consistent, and is not sensitive to image content, distortion type, or subject identity; such PIPs can be generated at very low cost. The proposed BIQA method is one of learning to rank. We first formulate the problem of learning the mapping from the image features to the preference label as one of classification. In particular, we investigate the utilization of a multiple kernel learning algorithm based on group lasso (MKLGL) to provide a solution. A simple but effective strategy to estimate perceptual image quality scores is then presented. Experiments show that the proposed BIQA method is highly effective and achieves comparable performance to state-of-the-art BIQA algorithms. Moreover, the proposed method can be easily extended to new distortion categories.
Introduction
Blind Image Quality Assessment (BIQA) is the task of predicting how a human would perceive the quality of an image without a "perfect" reference for comparison. Traditionally, this has been approached as a regression problem, where models are trained to predict a Mean Opinion Score (MOS) on a scale (e.g., 0-100). However, MOS-based training faces significant hurdles: human scoring is often imprecise, biased by image content (e.g., preferring a beautiful sunset over a plain wall despite technical distortions), and inconsistent across different observers or sessions. Furthermore, scaling these databases is difficult because scoring scales from different experiments rarely align.
The Learning to Rank Framework
This paper proposes a paradigm shift by treating BIQA as a Learning to Rank problem. Instead of asking "How good is this image?", the model is trained to answer "Is Image A better than Image B?"
Preference Image Pairs (PIPs)
The foundation of this approach is the Preference Image Pair (PIP). Training data consists of pairs of images with a binary label: +1 if image A is superior to B, and -1 otherwise. These relative labels are much more consistent across observers than absolute scores and can be generated at a significantly lower cost.
To ensure high-quality training data, the authors generate pairs from existing databases only when the quality difference between two images exceeds a specific threshold ($T$), filtering out ambiguous "too close to call" cases.
Feature Fusion and Multiple Kernel Learning
To represent image degradation, the model fuses three types of Natural Scene Statistics (NSS) features: BLIINDS-II, BRISQUE-L, and SRNSS. This results in an 84-dimensional vector capturing various aspects of technical quality, from wavelet coefficients to DCT domain statistics.
To process these features, the authors utilize a Multiple Kernel Learning algorithm based on Group Lasso (MKLGL). This allows the model to automatically weight different portions of the feature vector—representing different scales or properties—using a suite of 45 different Gaussian kernels.
Quality Prediction and Results
While the model learns to rank, it can still estimate absolute quality scores through a "gain" strategy. A test image is compared against the training set, and its final score is derived from its cumulative "wins" relative to the training images, mapped onto a standard 0-100 scale.
Key Findings
- State-of-the-Art Performance: The method achieved Spearman Rank Correlation Coefficient (SRCC) values comparable to or exceeding top-tier models like BRISQUE and CORNIA across major databases (LIVE, TID2013, CSIQ, and LIVEMD).
- Database Independence: Because the model learns relative relationships rather than arbitrary scales, a model trained on one database performs remarkably well on others without needing complex realignment.
- Scalability: Adding a new distortion type to the model requires only a few hundred preference pairs, rather than a full-scale, expensive subjective study.
Conclusion
By leveraging the human ability to compare rather than measure, this Learning to Rank framework provides a more robust and scalable solution for BIQA. Future work may explore more efficient active learning strategies to select image pairs and the integration of deep learning-based features to further improve performance on complex, simultaneous distortions.
Summary updated