Uncertainty-Aware Blind Image Quality Assessment in the Laboratory and Wild (2005.13983v6)

Abstract: Performance of blind image quality assessment (BIQA) models has been significantly boosted by end-to-end optimization of feature engineering and quality regression. Nevertheless, due to the distributional shift between images simulated in the laboratory and captured in the wild, models trained on databases with synthetic distortions remain particularly weak at handling realistic distortions (and vice versa). To confront the cross-distortion-scenario challenge, we develop a \textit{unified} BIQA model and an approach of training it for both synthetic and realistic distortions. We first sample pairs of images from individual IQA databases, and compute a probability that the first image of each pair is of higher quality. We then employ the fidelity loss to optimize a deep neural network for BIQA over a large number of such image pairs. We also explicitly enforce a hinge constraint to regularize uncertainty estimation during optimization. Extensive experiments on six IQA databases show the promise of the learned method in blindly assessing image quality in the laboratory and wild. In addition, we demonstrate the universality of the proposed training strategy by using it to improve existing BIQA models.

Uncertainty-Aware Blind Image Quality Assessment in the Laboratory and Wild

The paper "Uncertainty-Aware Blind Image Quality Assessment in the Laboratory and Wild" addresses the significant challenge in blind image quality assessment (BIQA) - the distributional shift between images simulated in a controlled lab environment and those captured in natural, realistic scenarios. This shift poses a challenge for models trained in one context to perform well in the other. To tackle this, the authors present a unified BIQA model trained on both synthetic and realistic distortions, aiming to improve cross-distortion-scenario performance.

Overview

Traditionally, BIQA models have been optimized for either synthetic distortions (e.g., Gaussian blur) or realistic distortions (e.g., sensor noise). This separation typically limits their applicability across different scenarios. The proposed model employs a deep neural network (DNN) framework and is trained using a novel approach that models uncertainty directly from the variance in human opinions. The training process leverages a large number of image pairs and uses a fidelity loss function together with a hinge constraint, enabling the model to approximate human uncertainty during quality assessment.

Key Contributions

1. Unified Training Strategy: The authors propose a method to train BIQA models across multiple databases with varied distortion types. They avoid perceptual scale realignment issues by using pairwise samples within each database and calculating the probability that one image is of higher quality than another, which is derived from human-rated subjective scores and their variances.
2. Uncertainty Modeling: A key innovation of this work is the explicit modeling of human opinion uncertainty in BIQA. The model learns to replicate human tendencies to give consistent (low variance) ratings for extreme-quality images and exhibit higher uncertainty for mid-quality images.
3. Robustness Across Databases: This model shows promising results in assessing image quality on six distinct databases containing both synthetic and realistic distortions. The results indicate significant improvements compared to several state-of-the-art BIQA models, particularly in terms of cross-distortion-scenario generalization.
4. Relevance in Practical Applications: The enhanced capability of this model to handle varied distortion scenarios without additional subjective testing makes it particularly valuable for real-world applications where images originate from diverse sources.

Experimental Results

The model's performance was evaluated on six image quality assessment (IQA) databases, showcasing superior Spearman rank-order and Pearson linear correlation coefficients relative to both full-reference IQA measures and contemporary BIQA models. The authors further validate the model's generalization through the group maximum differentiation (gMAD) competition, illustrating its robustness in perceptual quality prediction without human annotations.

Implications and Future Work

The implications of this research are twofold. Practically, it provides a more reliable BIQA tool that can be leveraged across various imaging systems to consistently monitor and enhance image quality, regardless of distortion context. Theoretically, it advances the understanding of uncertainty modeling in machine learning, specifically in the domain of IQA.

For future research, extending this approach to video quality assessment and exploring its application to other perceptual tasks could offer additional insights and practical benefits. Further exploration of model architectures and training strategies could also drive continuous improvement in the field of BIQA, aligning model predictions ever closer to human perceptual standards.