Image Quality Assessment: Exploring Quality Awareness via Memory-driven Distortion Patterns Matching

Xuting Lan; Mingliang Zhou; Xuekai Wei; Jielu Yan; Yueting Huang; Huayan Pu; Jun Luo; Weijia Jia

cs.CVFeb 20, 2026

Image Quality Assessment: Exploring Quality Awareness via Memory-driven Distortion Patterns Matching

Authors: Xuting Lan, Mingliang Zhou, Xuekai Wei, Jielu Yan, Yueting Huang, Huayan Pu, Jun Luo, Weijia Jia

Abstract

Existing full-reference image quality assessment (FR-IQA) methods achieve high-precision evaluation by analysing feature differences between reference and distorted images. However, their performance is constrained by the quality of the reference image, which limits real-world applications where ideal reference sources are unavailable. Notably, the human visual system has the ability to accumulate visual memory, allowing image quality assessment on the basis of long-term memory storage. Inspired by this biological memory mechanism, we propose a memory-driven quality-aware framework (MQAF), which establishes a memory bank for storing distortion patterns and dynamically switches between dual-mode quality assessment strategies to reduce reliance on high-quality reference images. When reference images are available, MQAF obtains reference-guided quality scores by adaptively weighting reference information and comparing the distorted image with stored distortion patterns in the memory bank. When the reference image is absent, the framework relies on distortion patterns in the memory bank to infer image quality, enabling no-reference quality assessment (NR-IQA). The experimental results show that our method outperforms state-of-the-art approaches across multiple datasets while adapting to both no-reference and full-reference tasks.

Figures & tables

Fig. 1: Comparison of MQAF with existing IQA methods, highlighting the exceptional IQA ability of MQAF-R.

Fig. 2: The framework (MQAF) of our proposed method can adaptively select different modes based on the availability of reference images in the quality assessment process. When a reference image is available, the framework enters the reference mode, where the final quality score $Q$ is obtained by computing reference matching and memory matching. When no reference image is available, the framework switches to the no-reference mode, where the final quality score $Q$ is determined by matching with distortion patterns (memory units $\mathcal{V}$ ) stored in the memory bank.

Dataset	LIVE [ 36 ]	CSIQ [ 27 ]	TID2013 [ 37 ]	KADID-10K [ 38 ]	PIPAL [ 39 ]
Reference images	29	30	25	81	250
Distorted images	779	866	3,000	10,125	25,850
Distortion types	5	6	24	25	40
Image resolution	-	512 $\times$ 512	512 $\times$ 384	512 $\times$ 384	288 $\times$ 288
MOS range	[0, 100]	[0, 1]	[0, 9]	[1, 5]	[868, 1857]

TABLE I: Detailed information of the image quality assessment databases.

Type	Method	LIVE [ 36 ]		CSIQ [ 27 ]		TID2013 [ 37 ]		KADID-10K [ 38 ]
Type	Method	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC
FR	PSNR	0.791	0.807	0.781	0.801	0.663	0.686	0.667	0.672
	SSIM [ 24 ]	0.847	0.851	0.819	0.832	0.665	0.627	0.780	0.778
	MS-SSIM [ 25 ]	0.886	0.903	0.864	0.879	0.785	0.729	0.835	0.834
	VIF [ 26 ]	0.948	0.952	0.898	0.899	0.771	0.677	0.676	0.669
	MAD [ 27 ]	0.904	0.907	0.934	0.932	0.803	0.773	0.829	0.827

TABLE II: Performance comparison of the proposed MQAF algorithm against state-of-the-art FR-IQA and NR-IQA algorithms on benchmark datasets. The best, second, and third results are bolded and highlighted in red , green , and blue , respectively. Additionally, MQAF-R presents the performance of our method in reference mode, while MQAF-NR shows its performance in no-reference mode.

Methods	PIPAL Validation [ 39 ]
Methods	PLCC	SRCC
PSNR	0.289	0.255
SSIM [ 24 ]	0.409	0.363
MS-SSIM [ 25 ]	0.566	0.491
FSIM [ 28 ]	0.562	0.468
VSI [ 29 ]	0.515	0.450

TABLE III: SRCC results of MQAF-R on the PIPAL dataset. The best results are shown in bold .

Method	Training Set: KADID-10K
	LIVE		CSIQ		TID2013
	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC
PSNR	0.872	0.0.876	0.796	0.806	0.702	0.639
WaDIQaM-FR [ 20 ]	0.837	0.883	-	-	0.741	0.698
RADN [ 56 ]	0.878	0.905	-	-	0.796	0.747
PieAPP [ 43 ]	0.910	0.918	0.890	0.897	0.831	0.844

TABLE IV: Comparison of cross-dataset performance on KADID-10K [ 38 ] dataset. The best results are highlighted in bold .

Method	LIVE					CSIQ
Method	JP2K	JPEG	WN	GB	FF	WN	JPEG	JP2K	PN	GB	CC
PSNR	0.895	0.881	0.985	0.782	0.891	0.963	0.888	0.936	0.934	0.929	0.862
SSIM [ 24 ]	0.961	0.972	0.969	0.952	0.956	0.897	0.956	0.961	0.892	0.961	0.792
VIF [ 26 ]	0.969	0.984	0.985	0.972	0.965	0.958	0.971	0.967	0.951	0.975	0.935
GSMD [ 40 ]	0.968	0.973	0.974	0.957	0.942	0.944	0.963	0.965	0.939	0.959	0.935
FSIMc [ 28 ]	0.972	0.979	0.971	0.968	0.950	0.936	0.966	0.970	0.937	0.973	0.944

TABLE V: SROCC results for individual distortion types on the LIVE [ 36 ] and CSIQ [ 27 ] databases. The best result is shown in bold and the second-best result is underlined .

Dist. type	BRISQUE [ 59 ]	ILNIQE [ 60 ]	CORNIA [ 61 ]	HOSA [ 62 ]	WaDIQaM-NR [ 20 ]	DBCNN [ 19 ]	MetalQA [ 47 ]	† MQAF-NR (Ours)	† MQAF-R (Ours)
AGN	0.9356	0.8760	0.4465	0.7582	0.9080	0.7585	0.9473	0.9650	* 0.9694
ANC	0.8114	0.8159	0.1020	0.4670	0.8700	0.8423	0.9240	0.9450	0.9188
SCN	0.5457	0.9233	0.6697	0.6246	0.8802	0.7800	0.9534	0.9758	0.9726
MN	0.5852	0.5120	0.6096	0.5125	0.8065	0.4501	0.7277	0.8362	* 0.8831
HFN	0.8956	0.8685	0.8402	0.8285	0.9314	0.8192	0.9518	0.9509	* 0.9667
IN	0.6559	0.7551	0.3526	0.1889	0.8779	0.5798	0.8653	0.8384	0.8172

TABLE VI: SROCC value comparison in leave–one-distortion-out cross-validation on the TID2013 dataset. The best result is shown in bold , the second-best result is underlined , and † denotes the reference mode (MQAF-R) and no-reference mode model (MQAF-NR), whereas

*

indicates the best results under the MQAF-R model.

NAME	PieAPP [ 43 ]	LPIPS [ 17 ]	DISTS [ 15 ]	PDL [ 32 ]	WaDIQaM-FR [ 20 ]	A-DISTS [ 16 ]
MACs / G	30.25	20.04	20.40	25.40	20.04	20.04
Params / M	68.38	14.71	16.02	20.02	14.72	14.71
Runtime / min	29.74	1.98	3.66	11.66	15.43	3.10
NAME	AHIQ [ 12 ]	DeepWSD [ 18 ]	TOPIQ [ 14 ]	DeepDC [ 63 ]	MQAF-R w/VGG16	MQAF-R w/ResNet50
MACs / G	1689.91	25.40	7.80	19.05	30.70	8.27
Params / M	133.39	20.02	32.35	26.54	16.29	28.22

TABLE VII: Comparison of computational complexity on different methods and backbone networks.

Method	LIVE		CSIQ		TID2013		KADID-10K
Method	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC
Memory-driven (w/o)	0.980	0.977	0.964	0.970	0.931	0.937	0.950	0.950
MQAF-R (Ours)	0.982	0.980	0.977	0.979	0.964	0.966	0.964	0.963
MQAF-NR (Ours)	0.984	0.980	0.971	0.972	0.963	0.960	0.962	0.958

TABLE VIII: Results of ablation results on four benchmark datasets. The best result is shown in bold .

Memory Units	$\mathbf{\lambda}$	PLCC	SRCC
128	0.01	0.963	0.962
256	0.01	0.964	0.964
512	0.01	0.963	0.964
1024	0.01	0.962	0.962
2048	0.01	0.962	0.959

TABLE IX: Effect of the number of memory units on model performance on the TID2013 [ 37 ] dataset.

$\mathbf{\lambda}$	Memory Units	PLCC	SRCC
0	256	0.963	0.962
0.001	256	0.964	0.964
0.005	256	0.963	0.963
0.01	256	0.960	0.963
0.05	256	0.963	0.964
0.1	256	0.964	0.966

TABLE X: Sensitivity analysis of the memory loss in Equation ( 9 ) hyperparameter

\lambda

on the TID2013 [ 37 ] dataset.

Fig. 3: Scatter plot of prediction results for different FR-IQA methods and our method (MQAF-R) on the KADID-10K [ 38 ] dataset. The vertical axis represents the ground truth values, whereas the horizontal axis represents the predicted values.

Fig. 4: Comparison of gMAD competition results on the TID2013 dataset among our proposed MQAF-R method, PieAPP [ 43 ] , WaDIQaM-FR [ 20 ] , and DISTS [ 15 ] is presented.

Explore similar work

Pixel Perfect: Relational Image Quality Assessment with Spatially-Aware Distortions
May 4, 2026Fadeel Sher Khan, Long N. Le, Abhinau K. Venkataramanan +2Image Quality Assessment
The University of Texas at Austin Austin, TX, USA · Samsung Research America Plano, TX, USA
SHAMISA: SHAped Modeling of Implicit Structural Associations for Self-supervised No-Reference Image Quality Assessment
Mar 14, 2026Mahdi Naseri, Zhou WangImage Quality Assessment Self-Supervised Learning
Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, ON N2L 3G1, Canada
Preference-Guided Debiasing for No-Reference Enhancement Image Quality Assessment
Mar 20, 2026Shiqi Gao, Kang Fu, Zitong Xu +4Image Quality Assessment Debiased ML
Institute of Image Communication and Network Engineering, Shanghai Jiao Tong University, Shanghai 200240, China

May 4, 2026J/K move · Enter open · S save

Pixel Perfect: Relational Image Quality Assessment with Spatially-Aware Distortions

Fadeel Sher Khan, Long N. Le, Abhinau K. Venkataramanan, Seok-Jun Lee+1

The University of Texas at Austin Austin, TX, USA · Samsung Research America Plano, TX, USA

Image Quality Assessment

Open paper AskPDF arXiv

Traditional image quality assessment (IQA) methods rely on mean opinion scores (MOS), which are resource-intensive to collect and fail to provide interpretable, localized feedback on specific image distortions. We overcome these limitations by shifting from absolute quality prediction to a relational and directional assessment. Our approach utilizes a self-supervised synthetic distortion engine to generate training data, eliminating the need for manual annotation. A distortion prediction network is trained with an anti-symmetric objective to produce spatially-aware, disentangled maps that identify the type, intensity, and direction of distortions relative to a reference image. Subsequently, a scoring network is trained via contrastive learning on ordinally ranked image sets to predict a relational quality score. Our method provides a more granular and interpretable approach to IQA for the targeted optimization of image processing algorithms without requiring any human-labeled quality scores.

Abstract

Explore similar work

Pixel Perfect: Relational Image Quality Assessment with Spatially-Aware Distortions

SHAMISA: SHAped Modeling of Implicit Structural Associations for Self-supervised No-Reference Image Quality Assessment

Preference-Guided Debiasing for No-Reference Enhancement Image Quality Assessment