Technical Summary: MR-IQA

Problem Statement

Blind Image Quality Assessment (BIQA) aims to model human perceptual quality judgments without reference images. Current approaches primarily rely on two learning paradigms: regression, which calibrates absolute quality scores, and ranking, which recovers quality structures from ordinal relations. While recent works, particularly those utilizing Multimodal Large Language Models (MLLMs), often employ joint regression-ranking supervision to improve performance, the theoretical relationship between these two paradigms remains largely empirical. Existing methods typically balance these objectives through heuristic loss weights or reward designs rather than a shared optimization principle. Furthermore, regression-based methods are often sensitive to dataset-specific score anchors, while ranking-based methods may fail to capture the continuous relative distance between images, focusing only on preference direction.

Methodology: MR-IQA

The authors propose MR-IQA, a unified framework that reinterprets both regression and ranking through the lens of quality margins.

1. Theoretical Foundation: The Unified Margin View

The paper derives that both regression and ranking fundamentally optimize pairwise relational distances (quality margins), defined as $\Delta\mu_{ij} = \mu_i - \mu_j$, where $\mu$ represents the Mean Opinion Score (MOS).

• Regression View: Pointwise regression is shown to optimize margin errors ($\delta_{ij} = \Delta s_{ij} - \Delta\mu_{ij}$) coupled with a dataset-anchor term ($\bar{s} - \bar{\mu}$). The global prediction shift in regression acts as a restrictive anchor that can hinder cross-dataset generalization.
• Ranking View: Thurstone-style ranking, which models preference probabilities $P_{ij}$, is mathematically equivalent to fitting transformed margins. The optimization direction of fidelity-based ranking loss converges to matching predicted margins ($\Delta s_{ij}$) with human margins ($\Delta\mu_{ij}$).
• The Bridge: The paper proves that the standard BIQA metric, Pearson Linear Correlation Coefficient (PLCC), is theoretically equivalent to the cosine similarity between the vector of predicted margins and the vector of human margins. This establishes margin fitting as the shared optimization target for both paradigms.

2. Algorithm: Reinforcement Learning with Margin Rewards

MR-IQA instantiates this theory as a Reinforcement Learning (RL) algorithm using Group Relative Policy Optimization (GRPO).

• Sampling: For a group of $N$ images, the policy model samples $K$ quality-score completions per image.
• Margin Calculation: For each sampled completion $s^{(k)}_i$, the model computes the predicted margin against every other image in the group: $\Delta s^{(k)}_{ij} = s^{(k)}_i - \bar{s}_j$.
• Reward Design: The model calculates a scale-controlled margin error $z^{(k)}_{ij}$ comparing the predicted margin to the ground-truth MOS margin ($\Delta\mu_{ij}$). Two scale options are explored:
  • Raw: $\tau_{ij} = 1$ (absolute mismatch).
  • Uncertainty-normalized: $\tau_{ij} = \sqrt{\sigma_i^2 + \sigma_j^2}$ (relative to inter-rater variance).
• Reward Conversion: The error is converted into a Gaussian or Laplace reward ($r_{margin}$). The final image-level reward aggregates pairwise rewards across the group, normalized by the group's mean and standard deviation to compute the advantage for policy updates.
• Training: The framework is trained on the KonIQ-10k dataset using Qwen-family MLLMs as backbones, optimizing directly for margin alignment without explicit pointwise score regression or discrete ranking labels.

Key Contributions

1. Theoretical Unification: The paper bridges regression and ranking by identifying quality margin as their common underlying quantity. It provides a theoretical derivation showing that both paradigms are essentially margin-oriented optimizations, with regression including an additional dataset-anchor term.
2. MR-IQA Framework: The authors instantiate this insight into MR-IQA, a scale-controlled RL algorithm. Unlike previous methods that combine regression and ranking losses, MR-IQA directly optimizes pairwise margin errors as policy rewards, explicitly modeling the continuous relational quality structure.
3. Empirical Validation: Through controlled experiments across six BIQA benchmarks (including in-distribution and out-of-distribution datasets), MR-IQA demonstrates competitive general performance. It achieves the strongest average PLCC and SRCC among RL-based methods, outperforming both pure regression and pure ranking baselines under matched settings.

Experimental Results

• Performance: On six benchmarks (KonIQ-10k, SPAQ, LIVE-Challenge, AGIQA-3K, KADID-10k, CSIQ), MR-IQA achieves an average PLCC of 0.831 and SRCC of 0.810. This is comparable to the strong SFT-based model DeQA (0.838/0.813) and significantly outperforms other RL-based methods like Q-Insight (regression) and VQ-R1 (ranking).
• Ablation Studies:
  • Reward Function: The L2 estimator with raw margin scale ($\tau_{ij}=1$) yielded the best results. Surprisingly, normalizing by human inter-rater variance did not consistently improve performance, suggesting that human variance is not always a reliable training-time scale for models.
  • Backbone Stability: The performance gains of MR-IQA over regression and ranking baselines were consistent across different MLLM backbones (Qwen3-VL-2B/4B and Qwen2.5-VL-7B), indicating the method's robustness is not tied to a specific model scale.
• Case Studies: Qualitative analysis shows that MR-IQA aligns visual evidence with MOS margins more robustly across distributions compared to baselines, which often exhibit perceptual overfitting or fail to capture relative distances accurately.

Significance and Claims

The paper claims that MR-IQA provides a new theoretical insight into unifying regression and ranking, moving beyond empirical combinations to a shared optimization principle based on quality structure.

• Theoretical Basis: It offers a foundation for understanding BIQA by showing that score calibration and ordinal comparison are unified through relational quality structure (margins).
• Design Principle: The work transforms the empirical complementarity between score calibration and ordinal comparison into an explicit optimization principle.
• Generalizability: The authors suggest this margin-based view extends beyond BIQA to other quality and ranking tasks, such as aesthetic assessment, video quality assessment, and learning-to-rank.
• Limitations: The authors modestly note that inter-rater variance normalization did not consistently help, and the method requires sufficient pair coverage, which may be noisy on small datasets. They also note that the current experiments focus on Qwen-family MLLMs.