Boosting Image Quality Assessment Performance: Unsupervised Score Fusion by Deep Maximum a Posteriori Estimation

Problem
This paper addresses the limitations of existing Image Quality Assessment (IQA) models, which often exhibit biases towards specific image content or distortion types due to their design principles. The authors identify a gap in the literature regarding effective methods for fusing scores from multiple IQA models to enhance overall performance. This work is presented as a preprint and has not yet undergone peer review.

Method
The authors propose a novel framework for unsupervised IQA score fusion utilizing deep Maximum a Posteriori (MAP) estimation. The architecture leverages a probabilistic approach to model the uncertainty associated with individual IQA scores, allowing for fine-grained uncertainty estimation at the score level. The training process involves optimizing the MAP estimation to derive a fused score that maximizes the posterior probability of the observed data given the model parameters. The authors do not disclose specific details regarding the dataset used for training or the computational resources required, but they emphasize the unsupervised nature of their approach, which does not rely on labeled data.

Results
The proposed method demonstrates significant improvements over individual IQA models and existing score fusion techniques. The authors report that their model outperforms state-of-the-art baselines on standard IQA benchmarks, although specific numerical results and effect sizes are not detailed in the abstract. The model’s ability to reject “bad” models during the fusion process is highlighted as a key advantage, suggesting that it can dynamically adjust the contribution of each model based on its reliability.

Limitations
The authors acknowledge that their approach may still be sensitive to the selection of the individual IQA models included in the fusion process. They do not address potential limitations related to the generalizability of their method across diverse image types or distortions, nor do they discuss the computational efficiency of their framework in practical applications. Additionally, the lack of labeled data for training may limit the model’s performance in scenarios where high-quality ground truth is available.

Why it matters
This work has significant implications for the field of image quality assessment, particularly in applications where accurate quality predictions are critical, such as in image compression, transmission, and enhancement. By providing a robust framework for score fusion, the authors pave the way for future research to explore more sophisticated ensemble methods that can leverage the strengths of various IQA models. The ability to dynamically reject less reliable models could lead to more resilient and adaptable IQA systems, enhancing the overall reliability of perceptual quality assessments in real-world scenarios.

Authors: Zhongling Wang, Raymond Zhou, Shahrukh Athar, Wenbo Yang, Zhou Wang
Source: arXiv:2605.30269
URL: https://arxiv.org/abs/2605.30269v1