HIFC-IQA: Train-Free Cross-Domain Image Quality Assessment via Dual-Process Cognition
Organizations: Harbin Institute of Technology · Taobao & Tmall Group
Abstract
The gap between controlled synthetic distortions and complex real-world degradations introduces a critical domain shift challenge for No-Reference Image Quality Assessment. Although Unsupervised Domain Adaptation (UDA) approaches attempt to address this issue, they remain constrained by substantial retraining overhead and cumulative error propagation inherent in iterative pseudo-labeling. Inspired by the dual-process theory of human cognition for remarkable generalization ability, we propose Hierarchical Intuitive and Fuzzy Consensus (HIFC), a completely train-free Cross-Domain IQA framework. HIFC employs asymmetric perceptual decomposition to disentangle semantic content from distortion-sensitive representations. The final quality score is fused through the synergy of two complementary branches: a memory consensus branch that retrieves a stable Fuzzy Quality Centroid from a synthetic reference gallery, and an intuitive branch that captures aesthetic polarization via zero-shot vision-language alignment. Extensive experiments show that HIFC achieves state-of-the-art performance across multiple benchmarks, attaining the highest average SRCC and PLCC among competing UDA methods. Most notably, in the challenging synthetic-to-authentic cross-domain setting, HIFC achieves a substantial performance margin over competing methods, demonstrating remarkable robustness and generalization.
Figures & tables
| Methods | Train | LIVEC [ 14 ] | KonIQ-10k [ 18 ] | BID [ 10 ] | Average | ||||
|---|---|---|---|---|---|---|---|---|---|
| SRCC | PLCC | SRCC | PLCC | SRCC | PLCC | SRCC | PLCC | ||
| RankIQA [ 33 ] | Yes | 0.491 | 0.495 | 0.603 | 0.551 | 0.510 | 0.367 | 0.535 | 0.471 |
| DBCNN [ 59 ] | Yes | 0.572 | 0.589 | 0.639 | 0.618 | 0.620 | 0.609 | 0.613 | 0.606 |
| HyperIQA [ 47 ] | Yes | 0.490 | 0.487 | 0.545 | 0.556 | 0.379 | 0.282 | 0.472 | 0.442 |
| MUSIQ [ 23 ] | Yes | 0.517 | 0.524 | 0.554 | 0.573 | 0.575 | 0.600 | 0.549 | 0.566 |
| VCRNet [ 37 ] | Yes | 0.561 | 0.548 | 0.517 | 0.525 | 0.542 | 0.545 | 0.540 | 0.540 |
| Distortion Type | DGQA [ 27 ] | Ours | ||
|---|---|---|---|---|
| SRCC | PLCC | SRCC | PLCC | |
| Traditional SR | 0.4897 | 0.4567 | 0.4981 | 0.4939 |
| PSNR-originated SR | 0.5419 | 0.5404 | 0.5627 | 0.5639 |
| SR with kernel mismatch | 0.5810 | 0.5956 | 0.5892 | 0.5981 |
| GAN-based SR | 0.1629 | 0.1353 | 0.3020 | 0.3010 |
| Denoising | 0.5393 | 0.5279 | 0.6172 | 0.5912 |
| Methods | Training | LIVE | CSIQ | TID2013 | |||
|---|---|---|---|---|---|---|---|
| SRCC | PLCC | SRCC | PLCC | SRCC | PLCC | ||
| DGQA | 0.9173 | 0.8988 | 0.7965 | 0.8017 | 0.7077 | 0.7220 | |
| SynDR-IQA | 0.9258 | 0.9014 | 0.8069 | 0.8092 | 0.7147 | 0.7328 | |
| CLIP-IQA | 0.8769 | 0.8703 | 0.7664 | 0.8027 | 0.6688 | 0.6873 | |
| Ours | 0.9036 | 0.8913 | 0.8113 | 0.8359 | 0.7041 | 0.6794 | |
| Method | IQA train | LIVEC | KonIQ-10k | BID |
|---|---|---|---|---|
| DGQA (ViT-B/32) | Yes | 0.712/0.721 | 0.697/0.701 | 0.723/0.731 |
| SynDR-IQA (ViT-B/32) | Yes | 0.732/0.736 | 0.752/0.748 | 0.745/0.754 |
| HIFC-IQA | No | 0.778/0.780 | 0.765/0.759 | 0.744/0.749 |
| KADID-10k subset | LIVEC | KonIQ-10k |
|---|---|---|
| 10% | 0.734/0.746 | 0.729/0.727 |
| 25% | 0.741/0.752 | 0.752/0.750 |
| 50% | 0.756/0.766 | 0.755/0.755 |
| 100% | 0.778/0.780 | 0.765/0.759 |
| Distortion diversity | LIVEC | KonIQ-10k |
|---|---|---|
| 5 types | 0.745/0.739 | 0.604/0.644 |
| 10 types | 0.767/0.764 | 0.612/0.647 |
| 20 types | 0.735/0.746 | 0.602/0.642 |
| Full diversity | 0.735/0.745 | 0.602/0.642 |
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
| Models | LIVEC | KONIQ | ||
|---|---|---|---|---|
| SRCC | PLCC | SRCC | PLCC | |
| RN50 | 0.700 | 0.705 | 0.693 | 0.707 |
| B-32 | 0.778 | 0.780 | 0.765 | 0.759 |
| B-16 | 0.718 | 0.750 | 0.770 | 0.756 |
| L-14 | 0.710 | 0.722 | 0.731 | 0.750 |
| Dataset | TOPK | Mean | Std | PLCC | SRCC | |
|---|---|---|---|---|---|---|
| KONIQ | 4 | 400 | 0.893 | 0.006 | 0.783 | 0.774 |
| LIVEC | 12 | 400 | 0.892 | 0.008 | 0.784 | 0.785 |
| BID | 16 | 700 | 0.903 | 0.009 | 0.758 | 0.764 |
| PIPAL | 4 | 500 | 0.900 | 0.006 | 0.508 | 0.513 |
| Method | Train | Inference Time(ms) | SRCC | PLCC |
|---|---|---|---|---|
| HyperIQA | 0.490 | 0.487 | ||
| DBCNN | 0.572 | 0.589 | ||
| CLIP-IQA | 0.612 | 0.594 | ||
| StyleAM | 0.584 | 0.561 | ||
| Ours | 0.778 | 0.780 |
| Metric | Value | Description |
|---|---|---|
| Backbone | CLIP (ViT-B/32) | Frozen, no training required |
| Parameters | 88 MB | Same as CLIP |
| FLOPs | 32.1G | Dual-branch (parallel) |
| GPU Memory | 1118 MB | During inference |
| Gallery Size | 10,125 | KADID-10k |
| Gallery Memory | 39 MB | Stored features (linear in ) |
| Reference | Method | LIVEC | KONIQ | ||
|---|---|---|---|---|---|
| SRCC | PLCC | SRCC | PLCC | ||
| NONE | CLIP-IQA | 0.612 | 0.594 | 0.695 | 0.717 |
| LIVEC | TReS | - | - | 0.733 | 0.739 |
| DBCNN | - | - | 0.754 | 0.759 | |
| Ours | - | - | 0.755 | 0.765 | |
| KONIQ | TReS | 0.768 | 0.775 | - | - |
| Ratio | KADID | KONIQ | SPAQ | |||
|---|---|---|---|---|---|---|
| SRCC | PLCC | SRCC | PLCC | SRCC | PLCC | |
| 9:1 | 0.841 | 0.852 | 0.822 | 0.836 | 0.880 | 0.882 |
| 8:2 | 0.831 | 0.836 | 0.826 | 0.839 | 0.882 | 0.883 |
| 5:5 | 0.809 | 0.813 | 0.825 | 0.829 | 0.873 | 0.873 |
| 2:8 | 0.778 | 0.786 | 0.809 | 0.815 | 0.860 | 0.861 |
| 1:9 | 0.743 | 0.758 | 0.801 | 0.803 | 0.847 | 0.843 |
| Dataset | Top-1 | Top-2 | Top-3 | Top-4 | Top-5 |
|---|---|---|---|---|---|
| LIVEC | JPEG 2000 | Motion Blur | JPEG | Darken | Pixelate |
| KONIQ | Motion Blur | Gaussian Blur | JPEG 2000 | Denoise | Color Diffusion |
| BID | Denoise | Motion Blur | Gaussian Blur | Lens Blur | JPEG 2000 |
| PIPAL | Motion Blur | JPEG 2000 | JPEG | Gaussian Blur | Contrast |
| Method | AGIQA-3K | AGIQA-20K | SAUD | SPAQ |
|---|---|---|---|---|
| BRISQUE | 0.178 | 0.118 | 0.216 | 0.314 |
| DBCNN | 0.324 | 0.284 | 0.418 | 0.559 |
| HyperIQA | 0.748 | 0.516 | 0.497 | 0.688 |
| MetaIQA | 0.468 | 0.348 | 0.586 | 0.646 |
| UCDA | 0.681 | 0.255 | 0.627 | 0.699 |
| DGQA | 0.577 | 0.498 | 0.651 | 0.705 |
| Method | AVA | PARA |
|---|---|---|
| NIMA | 0.612 / 0.636 | 0.571/0.587 |
| BRISQUE | 0.031/0.036 | 0.078/0.129 |
| Q-Align | 0.521/0.541 | 0.626/0.618 |
| DeQA-Score | 0.572/0.589 | 0.680/0.703 |
| DGQA | -/- | 0.322/0.351 |
| CLIP-IQA | 0.320/0.317 | 0.535/0.566 |
| Train/Reference | Method | Train | KonIQ-10k | |
|---|---|---|---|---|
| SRCC | PLCC | |||
| LIVE | PromptIQA | ✓ | 0.660 | 0.698 |
| Ours | 0.696 | 0.707 | ||
| CSIQ | PromptIQA | ✓ | 0.619 | 0.629 |
| Ours | 0.693 | 0.714 | ||
| KADID-10k | PromptIQA | ✓ | 0.577 | 0.595 |
| Semantic input | LIVEC | KonIQ-10k | Preprocessing (ms) |
|---|---|---|---|
| Direct resize | 0.692/0.720 | 0.682/0.687 | 3.439 |
| Letterbox resize | 0.712/0.733 | 0.671/0.675 | 3.483 |
| Center crop | 0.728/0.745 | 0.707/0.717 | 3.488 |
| Five random crops, averaged | 0.726/0.742 | 0.716/0.719 | 2.026 |
| Five-crop average | 0.773/0.778 | 0.751/0.752 | 1.786 |
| Proposed stacking | 0.778/0.780 | 0.765/0.759 | 1.086 |
| Distortion input | Overall SRCC | Overall PLCC | Color/illumination SRCC |
|---|---|---|---|
| RGB SJP | 0.6499 | 0.6569 | 0.3759 |
| Grayscale SJP | 0.6405 | 0.6442 | 0.3932 |
| Luminance-only SJP | 0.6432 | 0.6483 | 0.3997 |
| Methods | Train | LIVEC | KonIQ-10k | BID | Additional Domains (SRCC) | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| SRCC | PLCC | SRCC | PLCC | SRCC | PLCC | AGIQA-3K | AVA | PARA | PIPAL | ||
| MACLIP | No | 0.679 | 0.703 | 0.738 | 0.738 | 0.592 | 0.547 | 0.753 | 0.298 | 0.504 | 0.379 |
| LIQE | Yes | 0.677 | 0.680 | 0.719 | 0.726 | 0.684 | 0.652 | 0.653 | 0.429 | 0.483 | 0.480 |
| Q-Align [ 54 ] | Yes | 0.702 | 0.744 | 0.668 | 0.665 | 0.576 | 0.651 | 0.733 | 0.521 | 0.626 | 0.403 |
| DeQA-Score | Yes | 0.743 | 0.775 | 0.677 | 0.703 | 0.702 | 0.743 | 0.745 | 0.572 | 0.680 | 0.472 |
| Qwen2.5-VL-7B | Yes | 0.761 | 0.760 | 0.754 | 0.810 | 0.711 | 0.725 | – | – | – | – |
| Method | Retrieved information | Prediction | Distinguishing mechanism |
|---|---|---|---|
| IQARAG [ 13 ] | Reference images and MOSs | LMM score generation | Visual in-context quality anchors |
| QualiRAG [ 3 ] | Metadata, localization, and quality descriptions | Evidence-grounded LMM response | Retrieval of auxiliary quality knowledge |
| Wang et al. [ 50 ] | Subjectively scored images | Hierarchical kNN and MOS aggregation | Separate semantic and distortion classification modules |
| HIFC-IQA | Synthetic images and MOSs | FQC and zero-shot intuition | Frozen dual-space alignment and adaptive fuzzy consensus |
Image Quality Assessment
Image Quality Assessment measures perceived image quality, distortion, and usefulness with computational metrics and human-aligned evaluation.
See all 185 papers