BiRQA: Bidirectional Robust Quality Assessment for Images

Aleksandr Gushchin; Dmitriy S. Vatolin; Anastasia Antsiferova

cs.CVFeb 23, 2026

BiRQA: Bidirectional Robust Quality Assessment for Images

Authors: Aleksandr Gushchin, Dmitriy S. Vatolin, Anastasia Antsiferova

Organizations: ISP RAS Research Center for Trusted Artificial Intelligence, Moscow, Russia · MSU Institute for Artificial Intelligence, Moscow, Russia · Lomonosov Moscow State University, Moscow, Russia · Innopolis University, Innopolis, Russia

Abstract

Full-Reference image quality assessment (FR IQA) is important for image compression, restoration and generative modeling, yet current neural metrics remain slow and vulnerable to adversarial perturbations. We present BiRQA, a compact FR IQA metric model that processes four fast complementary features within a bidirectional multiscale pyramid. A bottom-up attention module injects fine-scale cues into coarse levels through an uncertainty-aware gate, while a top-down cross-gating block routes semantic context back to high resolution. To enhance robustness, we introduce Anchored Adversarial Training, a theoretically grounded strategy that uses clean "anchor" samples and a ranking loss to bound pointwise prediction error under attacks. On five public FR IQA benchmarks BiRQA outperforms or matches the previous state of the art (SOTA) while running ~3x faster than previous SOTA models. Under unseen white-box attacks it lifts SROCC from 0.30-0.57 to 0.60-0.84 on KADID-10k, demonstrating substantial robustness gains. To our knowledge, BiRQA is the only FR IQA model combining competitive accuracy with real-time throughput and strong adversarial resilience.

Figures & tables

Figure 1: Overall scheme of BiRQA. A reference–distorted pair yields four feature maps per pyramid level. Cross-Scale Residual Attention Module (CSRAM) and Spatial Cross-Gating Block (SCGB) allow the model to pass information in both directions between scales. A Reliability-Aware Head (GeM + dual MLPs) estimates per-level impact and reliability.

Figure 2: Scheme of the Cross-Scale Residual Attention Module (CSRAM) that lifts high-resolution cues to the next scale and injects them via uncertainty-aware gated residuals (strength $\alpha$ , confidence $\rho$ , and residual $R$ ) to refine the lower-resolution features.

Dataset	Resolution	Ref. Images	Dist. Images	Ratings
LIVE	$768\times 512$	29	779	25k
CSIQ	$512\times 512$	30	866	5k
TID2013	$512\times 384$	25	3,000	524k
KADID-10k	$512\times 384$	81	10,125	30.4k
PIPAL	$288\times 288$	250	23,200	1.13M
PieAPP	$256\times 256$	200	20,280	1M+

Table 1: Full-Reference IQA datasets used in our experiments.

Method	Trained on KADID-10k						Trained on PIPAL
Method	LIVE		CSIQ		TID2013		LIVE		CSIQ		TID2013
WaDIQaM-FR ( Bosse et al., 2017 )	0.940	0.947	0.901	0.909	0.834	0.831	0.895	0.899	0.834	0.822	0.786	0.739
PieAPP ( Prashnani et al., 2018 )	0.908	0.919	0.877	0.892	0.859	0.876	—	—	—	—	—	—
LPIPS-VGG ( Zhang et al., 2018 )	0.934	0.932	0.896	0.876	0.749	0.670	0.901	0.893	0.857	0.858	0.790	0.760
DISTS ( Ding et al., 2020 )	0.954	0.954	0.928	0.929	0.855	0.830	0.906	0.915	0.862	0.859	0.803	0.765
AHIQ ( Lao et al., 2022 )	0.952	0.970	0.955	0.951	0.889	0.885	0.903	0.920	0.861	0.865	0.804	0.763

Table 2: Cross-dataset performance on benchmarks (PLCC and SROCC are reported in corresponding columns for each benchmark). The best values are bolded, and the second best are underlined.

Model	SROCC $\uparrow$					IR-Score $\uparrow$				Train time
	Clean	FGSM	C&W	AutoAttack	FACPA	FGSM	C&W	AutoAttack	FACPA	(min)
base LPIPS	0.893	0.542	0.260	0.239	0.496	—	—	—	—	46
R-LPIPS	0.858	0.570	0.327	0.266	0.515	0.541	0.403	0.385	0.507	123
AT-LPIPS	0.852	0.730	0.523	0.481	0.753	0.722	0.596	0.510	0.613	101
AAT-LPIPS (ours)	0.860	0.751	0.578	0.532	0.796	0.788	0.649	0.580	0.628	339
base TOPIQ	0.938	0.524	0.284	0.269	0.512	—	—	—	—	352

Table 3: Robustness of adversarially trained IQA models on KADID-10k. SROCC is evaluated at

\epsilon=8/255

; IR-Score uses

\epsilon\!\in\!\{2,4,8,10\}/255

. Bold numbers denote the best result for each model. All adversarial variants were trained only with PGD-10.

Figure 3: Computational efficiency (FPS) vs. Performance (PLCC) comparison on PDAP-HDDS dataset with image size of

1920\times 1080

pixels. Our model achieves comparable PLCC to SOTA method TOPIQ, while being

\sim 3

times faster and having a notably smaller number of parameters.

CSRAM	SCGB	RAH	PLCC	SROCC
✗	✗	✗	0.801	0.813
✓	✗	✗	0.907	0.911
✓	✓	✗	0.925	0.928
✓	✓	✓	0.938	0.942

Table 4: Ablation study for BiRQA model. The CSRAM and SCGB modules were replaced with cross-attention layers and element-wise multiplication. The Reliability-Aware Head (RAH) module was replaced with pooling and MLP.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: SROCC and inference FPS on KADID-10k for different feature sets. The chosen quartet lies on the Pareto frontier, offering the best accuracy. FPS was measured on images with $1920\times 1080$ resolution.

Feature	Description / Key Benefit
SSIM	Structural Similarity Index compares luminance, contrast and structural components between a reference and a test image, yielding a single score that tracks perceived quality with high correlation to the human visual system (HVS).
Informational Map	Generates a spatial weighting map based on local information content (gradient magnitude), giving more influence to perceptually important, detail-rich regions.
Color Difference	Computes perceptual color difference in YCbCr color space, making the metric sensitive to chromatic distortions that luminance-only measures may miss.
LBP	Local Binary Patterns encode micro-texture by thresholding each neighborhood against its centre pixel; the histogram is gray-scale and rotation robust, providing a compact descriptor of fine texture changes.
Gabor Filters	A bank of Gabor kernels isolates edge and texture energy in specific frequency-orientation bands, capturing blur, ringing and other anisotropic artifacts.
Wavelet Transform	Discrete wavelet decomposition splits the image into multi-resolution sub-bands; analyzing coefficients across scales localizes blur or compression artifacts while preserving both spatial and frequency information.

Appendix

Table 5: Description of evaluated IQA features

Method	LIVE		CSIQ		TID2013		PieAPP		PIPAL
Method	PLCC	SROCC	PLCC	SROCC	PLCC	SROCC	PLCC	SROCC	PLCC	SROCC
PSNR	0.865	0.873	0.819	0.810	0.677	0.687	0.135	0.219	0.277	0.249
SSIM (2004)	0.937	0.948	0.852	0.865	0.777	0.727	0.245	0.316	0.391	0.361
MS-SSIM (2003)	0.940	0.951	0.889	0.906	0.830	0.786	0.051	0.321	0.163	0.369
VIF (2006)	0.960	0.964	0.913	0.911	0.771	0.677	0.250	0.212	0.479	0.397
MAD (2010)	0.968	0.967	0.950	0.947	0.827	0.781	0.231	0.304	0.580	0.543

Appendix

Table 6: Quantitative comparison with related works on public FR benchmarks, including the traditional LIVE, CSIQ, TID2013 with MOS labels, and recent large-scale datasets PieAPP, PIPAL with 2AFC labels. The best and second results are bold and underlined, respectively, and “—” indicates the score is not available or not applicable.

Figure 5: (a): Convergence of the Anchored Ranking Loss over 1,000 iterations. (b): Comparison of bounds, provided by Theorem 1 with empirical values of maximum pointwise errors.

Figure 6: Robustness of FR IQA metrics to FGSM attack with different perturbation budgets evaluated on KADID-10k dataset. Solid lines show SROCC of vanilla models without adversarial training, dashed lines show SROCC of models trained with AAT.

Model		White-Box Attacks					Black-Box Attacks
	Clean	FGSM	C&W	AutoAttack	FACPA	Zhang et al.	SquareAttack	Parsimonious
base LPIPS	0.742	0.260	0.102	0.135	0.415	0.301	0.511	0.513
R-LPIPS	0.728	0.487	0.306	0.298	0.471	0.375	0.567	0.533
AT-LPIPS	0.729	0.495	0.440	0.412	0.589	0.436	0.615	0.586
AAT-LPIPS (ours)	0.736	0.520	0.486	0.459	0.632	0.494	0.643	0.618
base TOPIQ	0.784	0.358	0.320	0.341	0.496	0.416	0.542	0.552

Appendix

Table 7: Accuracy on 2AFC BAPPS test set of different adversarial training techniques, which were applied to LPIPS, TOPIQ and BiRQA models. The best results for each model are bolded. PGD-10 with

\epsilon=8/255

was used during training.

	PSNR	SSIM	MAD	WaDIQaM	LPIPS	DISTS	AHIQ	TOPIQ	BiRQA (ours)
PSNR	—	—	—	—	—	—	—	—	—
SSIM	$0.103$ $\pm 0.003$	—	—	—	—	—	—	—	—
MAD	$0.281$ $\pm 0.005$	$0.178$ $\pm 0.001$	—	—	—	—	—	—	—
WaDIQaM	$0.292$ $\pm 0.007$	$0.189$ $\pm 0.001$	$0.011$ $\pm 0.001$	—	—	—	—	—	—
LPIPS	$0.336$ $\pm 0.007$	$0.232$ $\pm 0.001$	$0.054$ $\pm 0.003$	$0.044$ $\pm 0.001$	—	—	—	—	—
DISTS	$0.399$ $\pm 0.006$	$0.295$ $\pm 0.001$	$0.117$ $\pm 0.006$	$0.107$ $\pm 0.004$	$0.063$ $\pm 0.002$	—	—	—	—

Appendix

Table 8: Pairwise

\Delta

SROCC (±95% CI) on PIPAL (N = 23,000 distorted images) for 1,000 paired nonparametric bootstrap resamples (image pairs drawn with replacement; MOS and predictions kept aligned). Positive values favor the row metric; negative values favor the column metric. Only the lower triangle is shown; “—” indicates the symmetric counterpart.

Model		White-Box Attacks					Black-Box Attacks
	Clean	FGSM	C&W	AutoAttack	FACPA	Zhang et al.	SquareAttack	Parsimonious
base BiRQA	0.794	0.405	0.350	0.376	0.521	0.462	0.571	0.574
base BIRQA + Random Flip	0.751	0.471	0.467	0.422	0.569	0.570	0.606	0.608
base BIRQA + Random Rotate	0.720	0.446	0.431	0.390	0.534	0.512	0.541	0.553
base BIRQA + DiffPure	0.741	0.522	0.498	0.452	0.619	0.603	0.712	0.698
AAT-BiRQA (ours)	0.782	0.594	0.561	0.580	0.637	0.610	0.690	0.703

Appendix

Table 9: Accuracy on 2AFC BAPPS test set for AAT-BiRQA compared to some adversarial purification methods. The best results are bolded. PGD-10 with

\epsilon=8/255

was used during training of AAT-BiRQA.

Parameter	Value	Parameter	Value
Vanilla (clean) training		Anchored Adversarial Training (AAT)
Optimizer	Adam	Inner attacker	PGD-10
Adam $(\beta_{1},\beta_{2})$	0.9/0.999	PGD step size	$2/255$
Batch size	32	PGD norm type	$\ell_{\infty}$
Epochs	2500	Perturbation budget $\epsilon$	train: $8/255$ ; test: $\{2,4,8,10\}/255$
Learning rate	$10^{-4}$	Anchor spacing $\lambda$	0.5

Appendix

Table 10: List of hyper-parameters and description of experimental set-up for BiRQA.

Explore similar work

Learning Where to Look and How to Judge: Resolution-agnostic Image Quality Assessment with Quality-aware Saliency
Aug 3, 2026Hakan Emre Gedik, Shashank Gupta, Alan BovikImage Quality Assessment
The University of Texas at Austin · University of Colorado Boulder
MR-IQA-2: Faithful Image Quality Reflection via Fine-Grained Credit Assignment
Aug 19, 2026Yuan li, Youyuan Lin, Chenhui Chu +1Image Quality Assessment Faithfulness in VLMs
Graduate School of Informatics, Kyoto University
Low-Rank Adaptation of Frozen Vision-Language Models for Blind Image Quality Assessment
Jun 1, 2026Bishr Omer Adam, Xu LiVLM Adaptation Low-Rank Adaptation
School of Electronics and Information, Northwestern Polytechnical University, Xi’an, China

Aug 3, 2026J/K move · Enter open · S save

Learning Where to Look and How to Judge: Resolution-agnostic Image Quality Assessment with Quality-aware Saliency

Hakan Emre Gedik, Shashank Gupta, Alan Bovik

The University of Texas at Austin · University of Colorado Boulder

Image Quality Assessment

Open paper AskPDF arXiv

No-reference image quality assessment (NR IQA) has recently benefited from deep and multimodal models, yet many SOTA systems still violate at least one basic requirement: they either discard critical quality cues via aggressive resizing, fail to generalize across resolutions, cannot be jointly trained on heterogeneous IQA datasets with mismatched MOS scales, or require prohibitive computation. We present ReLIQS, a model for Resolution-agnostic Learning for Image Quality with Saliency, which is resolution-agnostic, preserves original-resolution quality cues, learns from multiple subjective studies, and remains computationally efficient and budget-adaptive. ReLIQS is a CLIP-based multiscale patch-driven architecture that learns both where to look and how to judge quality. Fixed-size patches are sampled across multiple resolutions, including the original resolution, and encoded with a CLIP vision backbone. A lightweight Perceptual Importance Estimator then predicts IQA-specific importance maps to select a small set of informative patches, and a Latent Quality Axis Module aggregates their embeddings into a single image-level score. Across authentic, synthetic, and AIGC benchmarks spanning diverse resolutions and distortions, ReLIQS generalizes better than strong CNN-, CLIP-, and MLLM-based baselines with matching or reduced computational cost.

Abstract

Explore similar work

Learning Where to Look and How to Judge: Resolution-agnostic Image Quality Assessment with Quality-aware Saliency

MR-IQA-2: Faithful Image Quality Reflection via Fine-Grained Credit Assignment

Low-Rank Adaptation of Frozen Vision-Language Models for Blind Image Quality Assessment