Abstract
Blind Image Quality Assessment (BIQA) models often struggle when deployed on datasets with different distortion types than those seen during training. This paper proposes a novel Source-free Unsupervised Domain Adaptation (SFUDA) method that adapts a pre-trained model to a target domain without requiring access to the original source data. By reframing BIQA as a rating distribution prediction task and enforcing human-centric Gaussian regularization, the proposed approach achieves state-of-the-art performance across various cross-domain and continual learning scenarios.
Introduction
The goal of Blind Image Quality Assessment (BIQA) is to predict human-perceived image quality without a reference image. While deep learning models have achieved high accuracy, they are often brittle, failing to generalize from synthetic distortions (e.g., Gaussian blur) to authentic, real-world distortions. Standard Unsupervised Domain Adaptation (UDA) techniques mitigate this but typically require simultaneous access to both source and target data. In many practical scenarios, source data is unavailable due to privacy or storage constraints. This work addresses the more challenging Source-free UDA (SFUDA) problem, where only a pre-trained model and unlabeled target data are available.
Methodology
The authors introduce a three-pronged strategy for source-free adaptation:
1. From Regression to Distribution Prediction
Unlike traditional BIQA models that output a single scalar score, this method treats quality assessment as a discrete probability distribution prediction over multiple quality levels (e.g., 1 to 5). This shift allows the model to capture the inherent subjectivity and uncertainty in human ratings.
2. Self-Supervised Objectives and Gaussian Regularization
To guide the model on unlabeled target data, three loss functions are employed:
- Entropy Loss: Minimizes uncertainty for individual image predictions.
- Diversity Loss: Prevents the model from collapsing to a single output value across the dataset.
- Gaussian Regularization: This "human-centric" constraint forces the predicted distribution for each image to follow a Gaussian shape, mimicking the typical distribution of human opinions for a single stimulus.
3. Efficient Fine-Tuning via Batch Normalization
To prevent catastrophic forgetting and ensure fast adaptation, the method only updates the affine parameters of Batch Normalization (BN) layers. Domain-Specific BN (DSBN) is used to handle multiple target domains, effectively capturing domain-specific "styles" while preserving general feature extraction capabilities.
Experimental Results
The method was evaluated on five major IQA benchmarks, including synthetic (LIVE, CSIQ, KADID-10K) and authentic (BID, KonIQ-10K) distortions, covering 20 cross-domain scenarios.
- Performance Gains: The proposed method significantly outperformed non-adapted baselines and even surpassed some UDA methods that utilized source data. For instance, in the KonIQ-10K to LIVE adaptation, the Spearman correlation (SROCC) improved by 0.1311.
- Continual Learning: In sequential learning tasks (e.g., LIVE → CSIQ → BID), the model maintained high performance on previous domains without catastrophic forgetting, thanks to the modular nature of the DSBN layers.
Conclusion
This research demonstrates that IQA can be effectively adapted to new domains without source data by leveraging human-centric priors. By treating quality as a distribution and enforcing Gaussian constraints, the model aligns more closely with human perception. Future work will explore more sophisticated probabilistic models to further capture the nuances of human subjectivity in image quality assessment.
Summary updated