Self-Supervised BIQA Methods

• Self-supervised BIQA is a technique for blind image quality assessment that predicts perceptual quality without using pristine references or manual annotations.
• Key methodologies include collaborative autoencoding for content-distortion disentanglement, quality-aware contrastive learning, and pseudo-labeling with full-reference agents, yielding up to 15% SRCC improvements.
• Recent advancements demonstrate robust cross-domain performance and effective source-free domain adaptation, ensuring reliable quality predictions on authentic, in-the-wild distortions.

Self-supervised Blind Image Quality Assessment (BIQA) refers to the class of methods for perceptual image quality prediction without reference images and without requiring manual, subjectively annotated quality scores for pre-training. Recent advances systematically leverage architectural innovations, synthetic data, and task-specific self-supervised (SSL) objectives to overcome the prohibitive costs and subjectivity of collecting large labeled IQA datasets, while also addressing cross-domain generalization. This entry surveys representative techniques, architectural paradigms, and experimental findings from major self-supervised BIQA approaches.

1. Problem Definition and Motivation

Blind Image Quality Assessment (BIQA) tasks require the prediction of perceptual quality (e.g., Mean Opinion Score, MOS) for a single image in the absence of a distortion-free reference. Conventional datasets annotated with MOS are limited in size and diversity, creating a bottleneck for data-hungry deep learning models (Zhou et al., 2023, Zhao et al., 2023). Empirical evidence demonstrates that deep models trained solely on human-annotated IQA data exhibit poor generalization to authentic, in-the-wild distortions and remain susceptible to domain shift (Wang et al., 2021, Liu et al., 2022).

Self-supervised BIQA seeks to sidestep this annotation bottleneck by:

• Exploiting the compositional nature of image degradations to synthesize training signals.
• Architecturally disentangling or suppressing nuisance variables (e.g., content vs. distortion).
• Transferring or ranking quality information using pseudo-labels from full-reference IQA (“agent”) models or properties of rating distributions.

These efforts aim to maintain high perceptual correlation (e.g., SRCC, PLCC) with ground-truth MOS, improve cross-database robustness, and minimize reliance on human opinion scores.

2. Collaborative Auto-Encoding and Feature Disentanglement

The “Collaborative Auto-encoding for Blind Image Quality Assessment” framework introduces a pair of collaborative autoencoders (COAE) to achieve architectural disentanglement of image content and distortion, forming the basis for downstream BIQA regression (Zhou et al., 2023).

Architectural Components

• Content Autoencoder (CAE): Trained exclusively on pristine images, the CAE encoder ($\mathbb{E}_c$) captures high-dimensional content codes $F_c \in \mathbb{R}^{256\times h\times w}$.
• Distortion Autoencoder (DAE): Trained on distorted images, the DAE encoder ($\mathbb{E}_d$) must reconstruct the input with side information from the CAE’s content code. Its low-capacity latent $f_d \in \mathbb{R}^{256}$, formed from Spatial Pyramid Pooling across four feature stages, is forced to encode only the distortion.

Feature Modulation

DAE’s decoder receives both the spatial content code $F_c$ and $f_d$. Via feature-modulating “Sub-Modulation Residual Blocks,” the model reconstructs the distorted image so that $f_d$ is maximally informative of the applied degradation.

Training Objective

Both autoencoders optimize a composite of $\ell_2$ reconstruction loss and perceptual similarity (LPIPS), but without explicit disentanglement regularization:

• $$L_{CAE} = \|I_{cin} - I_{cout}\|_F + LPIPS(I_{cin}, I_{cout})$$
• $$L_{DAE} = \|I_{din} - I_{dout}\|_F + LPIPS(I_{din}, I_{dout})$$

Quality Regression

Once CAE/DAE are pretrained, their encoders ($F_c \in \mathbb{R}^{256\times h\times w}$0) are frozen. For BIQA, features from both are pooled ($F_c \in \mathbb{R}^{256\times h\times w}$1), concatenated, and mapped to a quality score by a shallow MLP regressor fine-tuned on limited labeled MOS data.

This approach achieves strong within- and cross-database performance, outperforming decoupled or non-collaborative variants by up to 10-15% SRCC on representative BIQA benchmarks (Zhou et al., 2023).

3. Self-Supervised Pre-training via Quality-aware Contrastive Learning

The “Quality-aware Pre-trained Models for Blind Image Quality Assessment” approach adapts contrastive self-supervised learning (SSL) to BIQA by designing both view generation and losses sensitive to distortion, not merely semantics (Zhao et al., 2023).

Degradation Process and View Generation

A large-scale degradation space—∼$F_c \in \mathbb{R}^{256\times h\times w}$2 possible compositions—spans geometric, color, and texture-based image operations. For each unlabeled image, multiple degraded “views” are generated using random subsets and orders of degradation operators.

Patch-Level Contrastive Pretext Task

Within a degraded view, two spatially distinct patches (positive pair) share both content and distortion. Negatives include:

• Intra-image, cross-degradation (same content, different degradation).
• Inter-image pairs (content and likely degradation mismatch).

Quality-aware Contrastive Loss

The two-term loss: $F_c \in \mathbb{R}^{256\times h\times w}$3 with $F_c \in \mathbb{R}^{256\times h\times w}$4, $F_c \in \mathbb{R}^{256\times h\times w}$5, ensures that representations cluster by both content and perceptual quality, rather than semantics.

Empirical Outcomes

Networks pre-trained using this scheme and downstream fine-tuned on MOS data outperformed or matched transformer-based and hand-crafted BIQA methods across five real-world datasets (e.g., BID, CLIVE, KonIQ-10k, SPAQ, FLIVE), with up to 10% SRCC/PLCC gains over supervised-ImageNet-initialized backbones (Zhao et al., 2023).

4. Agent-driven and Opinion-free Self-supervision

Another direction leverages full-reference (FR) IQA algorithms as “agents” to provide pairwise pseudo-labels for large pools of synthetically distorted images (Wang et al., 2021).

Pseudo-labels from FR-IQA Agents

Large, diverse pairs of synthetically-distorted images are generated, and M=6 FR-IQA models assign pseudo-binary labels indicating which member of each pair is of higher quality. The method introduces agent reliability parameters ($F_c \in \mathbb{R}^{256\times h\times w}$6, $F_c \in \mathbb{R}^{256\times h\times w}$7) to mitigate label noise from imperfect agents.

Probabilistic Ranking and Domain Adaptation

A CNN-based BIQA model predicts both a Gaussian parameterized quality mean and uncertainty for each image; the learned probability that $F_c \in \mathbb{R}^{256\times h\times w}$8 is preferred over $F_c \in \mathbb{R}^{256\times h\times w}$9 is optimized via a Thurstone-V-based likelihood over all agent labels. An auxiliary Consensus Learning Classifier further regularizes optimization via consensus cross-entropy.

Unsupervised domain adaptation is incorporated via adversarial domain alignment and pixel-level mixup, reducing the synthetic-to-authentic domain gap without requiring real MOS during training.

Evaluation

This “opinion-free” framework achieves SRCC/PLCC ≥0.84 on SPAQ and ≥0.72 on KonIQ-10k when evaluated against human-annotated MOS—without using subjective labels during training (Wang et al., 2021).

5. Self-supervised Objectives for Source-free Domain Adaptation

Source-free unsupervised domain adaptation (SFUDA) for BIQA addresses the problem of transferring models to new, unlabeled target domains without access to source data (Liu et al., 2022).

Distributional Prediction

Instead of regressing a scalar, the model predicts a rating distribution $\mathbb{E}_d$0 over discrete quality levels $\mathbb{E}_d$1 (typically 1–5). During initial source training, ground-truth distributions ($\mathbb{E}_d$2) are drawn from truncated Gaussians centered at known MOS, and KL-divergence plus $\mathbb{E}_d$3-distance to expected value $\mathbb{E}_d$4 are minimized.

Self-supervised Losses for Adaptation

On the target domain, three self-supervised losses are minimized over model softmax predictions:

• Prediction Entropy Minimization: Drives output distributions toward confident (sharp) predictions.
• Batch Diversity Maximization: Ensures the model does not collapse to predicting the same rating for all images in a batch.
• Gaussian Regularization: Enforces consistency of softmax-predicted rating distributions with truncated Gaussians.

Only the BatchNorm affine parameters are adapted for the target domain, leveraging Domain-Specific Batch Normalization (DSBN).

Effectiveness

Experimental results indicate that applying all three self-supervised losses robustly mitigates cross-domain performance degradation. For example, when adapting from KADID-10K to KONIQ-10k, SROCC is improved by +0.084, and ablation studies reveal that only combinations of all three losses consistently improve generalization (Liu et al., 2022). The same DSBN-based adaptation strategy can be extended to continual learning across multiple target domains.

6. Quantitative Performance and Comparative Assessment

Representative self-supervised BIQA frameworks achieve state-of-the-art performance on both synthetic and authentic IQA benchmarks.

Method	Dataset	SRCC	PLCC	Protocol/Notes
COAE/“VISOR” (Zhou et al., 2023)	LIVE	0.973	0.978	80/20 split, within-database
	CSIQ	0.961	0.967
	TID2013	0.905	0.922
	KonIQ-10k	0.896	0.910
QPT (Zhao et al., 2023)	BID	0.8875	0.9109	Fine-tuned on labeled data
	KonIQ-10k	0.9271	0.9413
Opinion-free (Wang et al., 2021)	KonIQ-10k	0.717	0.740	No real labels during training
	SPAQ	0.838	0.844	No real labels during training
SFUDA (Liu et al., 2022)	KonIQ-10k	0.722	0.712	Source-free, after adaptation
	BID	0.595	0.598

Across methods, a common finding is that disentangling content from distortion, large and realistic synthetic degradation spaces, pairwise or distributional supervision, and domain adaptation are critical for robust BIQA, especially when scaling to unseen authentic distortions or target domains.

7. Ablation and Design Analysis

Comprehensive ablation studies elucidate the importance of self-supervised structures and objectives:

• The collaborative CAE/DAE architecture outperforms separate autoencoders by 10–15% SRCC, confirming the necessity of content-to-distortion information flow (Zhou et al., 2023).
• For QPT, removing either term in the two-part contrastive loss sharply degrades performance, as does restricting degradation diversity (Zhao et al., 2023).
• In SFUDA-style approaches, only the concurrent use of sharpness, diversity, and Gaussianity constraints leads to reliable adaptation gains; single/mixed losses only partially solve the domain gap (Liu et al., 2022).
• Adversarial and mixup domain adaptation, consensus learning regularization, and the use of agent reliabilities all measurably improve “opinion-free” BIQA (Wang et al., 2021).

A plausible implication is that further architectural or objective entanglement between distortion modeling and self-supervision will continue to characterize robust future BIQA developments.

---

References:

• "Collaborative Auto-encoding for Blind Image Quality Assessment" (Zhou et al., 2023)
• "Quality-aware Pre-trained Models for Blind Image Quality Assessment" (Zhao et al., 2023)
• "Source-free Unsupervised Domain Adaptation for Blind Image Quality Assessment" (Liu et al., 2022)
• "Learning from Synthetic Data for Opinion-free Blind Image Quality Assessment in the Wild" (Wang et al., 2021)