No-Reference Image Quality Assessment (NR-IQA) aims to estimate perceptual quality without access to a reference image of pristine quality. Learning an NR-IQA model faces a fundamental bottleneck: its need for a large number of costly human perceptual labels. We propose SHAMISA, a non-contrastive self-supervised framework that learns from unlabeled distorted images by leveraging explicitly structured relational supervision. Unlike prior methods that impose rigid, binary similarity constraints, SHAMISA introduces implicit structural associations, defined as soft, controllable relations that are both distortion-aware and content-sensitive, inferred from synthetic metadata and intrinsic feature structure. A key innovation is our compositional distortion engine, which generates an uncountable family of degradations from continuous parameter spaces, grouped so that only one distortion factor varies at a time. This enables fine-grained control over representational similarity during training: images with shared distortion patterns are pulled together in the embedding space, while severity variations produce structured, predictable shifts. We integrate these insights via dual-source relation graphs that encode both known degradation profiles and emergent structural affinities to guide the learning process throughout training. A convolutional encoder is trained under this supervision and then frozen for inference, with quality prediction performed by a linear regressor on its features. Extensive experiments on synthetic, authentic, and cross-dataset NR-IQA benchmarks demonstrate that SHAMISA achieves strong overall performance with improved cross-dataset generalization and robustness, all without human quality annotations or contrastive losses.
Figures & tables
Fig. 1: Overview of the proposed SHAMISA framework. (1) Self-supervised pre-training: pristine images are transformed by a compositional distortion engine to form a mini-batch X and distortion metadata; the encoder fθ and projector gψ produce representations H and embeddings Z used to construct dual-source relation graphs (metadata-driven and structurally intrinsic), which are aggregated by a multi-source aggregator Φ into G and optimize a graph-weighted SSL objective. (2) Downstream NR-IQA: we freeze fθ and train a lightweight regressor on top of H to predict quality scores.
Synthetic distortions
Authentic distortions
Method
Type
LIVE
CSIQ
TID2013
KADID-10K
FLIVE
SPAQ
Average
SRCC
PLCC
SRCC
PLCC
SRCC
PLCC
SRCC
PLCC
SRCC
PLCC
SRCC
PLCC
SRCC
PLCC
BRISQUE [ 20 ]
Handcrafted
0.939
0.935
0.746
0.829
0.604
0.694
0.528
0.567
0.288
0.373
0.809
0.817
0.652
0.703
NIQE [ 21 ]
Handcrafted
0.907
0.901
0.627
0.712
0.315
0.393
0.374
0.428
0.211
0.288
0.700
0.709
0.522
0.572
CORNIA [ 24 ]
Codebook
0.947
0.950
0.678
0.776
0.678
0.768
0.516
0.558
–
–
0.709
0.725
–
–
HOSA [ 25 ]
Codebook
0.946
0.950
0.741
0.823
0.735
0.815
0.618
0.653
–
–
0.846
0.852
–
–
TABLE I: Performance comparison on synthetic and in-the-wild NR-IQA datasets. We report SRCC and PLCC. Best and second-best per column are bold and underlined . Prior numbers are taken from their papers or public code where available [ 17 , 11 , 18 ] .
Training
Testing
HyperIQA
Su et al.
CONTRIQUE
Re-IQA
ARNIQA
SHAMISA
LIVE
CSIQ
0.744
0.777
0.803
0.795
0.904
0.909
LIVE
TID2013
0.541
0.561
0.640
0.588
0.697
0.700
LIVE
KADID-10K
0.492
0.506
0.699
0.557
0.764
0.694
CSIQ
LIVE
0.926
0.930
0.912
0.919
0.921
0.939
CSIQ
TID2013
0.541
0.550
0.570
0.575
0.721
0.729
CSIQ
KADID-10K
0.509
0.515
0.696
0.521
0.735
0.711
TABLE II: Cross-dataset SRCC on synthetic NR-IQA benchmarks. Best and second-best per row are bold and underlined . Baseline values are taken from prior reports in ARNIQA and Re-IQA [ 18 , 11 ] .
Fig. 2: gMAD on Waterloo Exploration (distorted pool): panels (a,b) use SHAMISA as the defender and panels (c,d) use ARNIQA as the defender. In each panel, the top and bottom images are the attacker-selected pair from the indicated defender-quality bin, with low-bin cases shown in (a,c) and high-bin cases shown in (b,d). Supplementary Table VII reports the corresponding top-1 attacker gaps.
Fig. 3: t-SNE visualization of image-level encoder representations on KADID-10K. We extract encoder representations H (pre-projector) and average across crops to obtain one representation per image, yielding 10,125 image-level points. Points are colored by the coarse distortion family using a shared fixed palette. Compared to ARNIQA, SHAMISA exhibits clearer coarse-family structure for several dominant degradation families, while both models show overlap among visually related families such as brightness and color manipulations. This visualization is a qualitative diagnostic and does not affect the quantitative evaluation protocol.
Fig. 4: Manifold visualization with UMAP [ 51 ] using encoder representations H extracted from 1,000 pristine KADIS images degraded with Gaussian blur and white noise. For each pristine image we generate 5 blur-only samples, 5 noise-only samples, and all 5×5 blur → noise compositions (35 variants per image, 35k total points). Point color is the weighted average of blur (red) and noise (yellow) according to their relative intensities; marker opacity increases with total severity.
Variant
SRCC g
PLCC g
SRCC a
PLCC a
SRCC s
PLCC s
A0: SHAMISA
0.8862
0.9042
0.7620
0.8040
0.9483
0.9543
(A) Aggregation & regularization
A1: Binary G
0.8767
0.8924
0.7475
0.7821
0.9374
0.9457
A3: No Φ (fixed linear mixture)
0.8759
0.8951
0.7479
0.7896
0.9361
0.9457
(B) Graph sources
B1: No Go (OT graph)
0.8600
0.8787
0.7661
0.8069
0.8995
0.9092
TABLE III: Main ablations . Median over 10 splits per dataset; then averaged. SRCC s /PLCC s average over synthetic {LIVE,CSIQ,TID2013,KADID\mbox−10K} , SRCC a /PLCC a over authentic {FLIVE,SPAQ} , and SRCC g /PLCC g denote the average over the six per-dataset median SRCC/PLCC values. Full ablation catalog is reported in Supplementary Table VIII.
Fig. 5: SRCC vs. training steps during SSL pre-training (mean with ± std error bars over 10 splits) across representative NR-IQA benchmarks. Curves typically rise early and then saturate, while the best checkpoint can vary by dataset since the SSL objective is not identical to downstream SRCC.
Fig. 6: One-at-a-time sweeps around the SHAMISA reference A0, reporting SRCC across six NR-IQA benchmarks in a single merged view. Columns correspond to (dh,dz) , (B,ξ) , (α,β) , (γ,η) , and (kn,K) , with a shared legend across all panels.
kNN graph: kn=31 ; metadata graph sparsifier: Kd=4096 ; ref-ref edge weight wrr=0.5766 ; OT graph: K=32 , τc=0.1 ; OT global sparsifier: Kg=8N .
Distortion engine
Per update: B=2 tiny-batches, R=3 references, C=4 composition groups, L=5 severity levels; crop size 224 ; max distortions per sample Md=7 .
Appendix
TABLE IV: Essential final hyperparameters used for SHAMISA pre-training. A single configuration is shared across all datasets (no dataset-specific tuning during SSL).
Dataset
Type
Pristine refs
Distortion types
Images
Label type
Split note
LIVE [ 41 ]
Synthetic
29
5
779
DMOS
Reference-disjoint
CSIQ [ 42 ]
Synthetic
30
6
866
DMOS
Reference-disjoint
TID2013 [ 43 ]
Synthetic
25
24
3000
MOS
Reference-disjoint
KADID-10K [ 9 ]
Synthetic
81
25
10,125
MOS
Reference-disjoint
FLIVE [ 29 ]
In-the-wild
–
–
39,808
MOS + patch scores
Official split
SPAQ [ 44 ]
In-the-wild
–
–
11,125
MOS
Random 70/10/20
Appendix
TABLE V: Dataset statistics for the NR-IQA benchmarks. Synthetic datasets are reported by reference count and distortion inventory, whereas authentic datasets are summarized by image count and label type.
Method
Type
LIVE
CSIQ
TID2013
KADID-10K
SRCC
PLCC
SRCC
PLCC
SRCC
PLCC
SRCC
PLCC
PSNR
Traditional
0.881
0.868
0.820
0.824
0.643
0.675
0.677
0.680
SSIM [ 52 ]
Traditional
0.921
0.911
0.854
0.835
0.642
0.698
0.641
0.633
FSIM [ 53 ]
Traditional
0.964
0.954
0.934
0.919
0.852
0.875
0.854
0.850
VSI [ 54 ]
Traditional
0.951
0.940
0.944
0.929
0.902
0.903
0.880
0.878
PieAPP [ 55 ]
Deep learning
0.915
0.905
0.900
0.881
0.877
0.850
0.869
0.869
Appendix
TABLE VI: Full-reference IQA comparison on synthetic distortion datasets. Reported rows are taken from prior work. SHAMISA-FR uses the same fixed-encoder evaluation protocol as our NR setting, and each value is the median over 10 random splits. Best and second-best values in each column are shown in bold and underlined.
Fig. 7: t-SNE colored by distortion severity on KADID-10K. Colors indicate severity levels 1 to 5 for ARNIQA (left) and SHAMISA (right).
Fig. 8: Subtype-colored t-SNE views for SHAMISA, including six KADID-10K distortion families.
Defender, Quality Level
Bin size
Top-1 attacker gap
ARNIQA, high
9488
8.2717
ARNIQA, low
9488
8.6782
SHAMISA, high
9488
4.1860
SHAMISA, low
9488
3.1356
Appendix
TABLE VII: gMAD summary on Waterloo Exploration (distorted pool) between SHAMISA and ARNIQA. Attacker gap denotes the attacker-side score separation of the selected pair within the defender-defined bin; lower is better.
Variant
SRCC g
PLCC g
SRCC a
PLCC a
SRCC s
PLCC s
A0: SHAMISA
0.8862
0.9042
0.7620
0.8040
0.9483
0.9543
(A) Graph weighting & regularization
A1: Binary G
0.8767
0.8924
0.7475
0.7821
0.9374
0.9457
A2: No Rgraph
0.8813
0.8981
0.7471
0.7908
0.9450
0.9497
A3: No Φ (fixed linear mixture)
0.8759
0.8951
0.7479
0.7896
0.9361
0.9457
A4: No stop-grad through Φ
0.8797
0.8980
0.7521
0.7950
0.9395
0.9469
Appendix
TABLE VIII: Full ablation catalog. For each dataset we run 10 random splits and report the median test SRCC/PLCC across splits, then average within groups: synthetic {LIVE,CSIQ,TID2013,KADID\mbox−10K} , authentic {FLIVE,SPAQ} , and SRCC g /PLCC g denote the average over the six per-dataset median SRCC/PLCC values.
Fig. 9: Feature decorrelation proxy ( corr ) over SSL pre-training. Both H and Z show a rapid early drop followed by stabilization, consistent with progressively reduced redundancy across feature dimensions.
Fig. 10: Effective dimensionality proxy ( rank ) for H over SSL pre-training. The encoder representation uses a substantial number of active dimensions throughout training and then stabilizes after the early phase.
Fig. 11: Invariance proxy ( inv ) over SSL pre-training. Curves generally decrease and then stabilize, consistent with paired views becoming closer in representation space as training proceeds.
HP
LIVE
CSIQ
TID2013
KADID-10K
FLIVE
SPAQ
Mean
dh
0.012
0.003
0.003
0.006
0.004
0.007
0.006
dz
0.006
0.002
0.002
0.000
0.003
0.007
0.003
B
0.002
0.001
0.003
0.005
0.001
0.001
0.002
kn
0.007
0.009
0.006
0.016
0.007
0.001
0.008
K
0.019
0.035
0.061
0.030
0.007
0.005
0.026
γ
0.017
0.045
0.089
0.070
0.033
0.001
0.042
Appendix
TABLE IX: Sensitivity summary for SRCC. For each hyperparameter, Δ is the largest minus smallest SRCC across the sweep. Larger Δ indicates higher sensitivity. The most sensitive and second most sensitive hyperparameters per dataset are shown in bold and underlined .
Current no-reference image quality assessment (NR-IQA) models for enhanced images often struggle to generalize, as they tend to overfit to the distinct patterns of specific enhancement algorithms rather than evaluating genuine perceptual quality. To address this issue, we propose a preference-guided debiasing framework for no-reference enhancement image quality assessment (EIQA). Specifically, we first learn a continuous enhancement-preference embedding space using supervised contrastive learning, where images generated by similar enhancement styles are encouraged to have closer representations. Based on this, we further estimate the enhancement-induced nuisance component contained in the raw quality representation and remove it before quality regression. In this way, the model is guided to focus on algorithm-invariant perceptual quality cues instead of enhancement-specific visual fingerprints. To facilitate stable optimization, we adopt a two-stage training strategy that first learns the enhancement-preference space and then performs debiased quality prediction. Extensive experiments on public EIQA benchmarks demonstrate that the proposed method effectively mitigates algorithm-induced representation bias and achieves superior robustness and cross-algorithm generalization compared with existing approaches.
Shiqi Gao, Kang Fu, Zitong Xu +4
Institute of Image Communication and Network Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
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.
Fadeel Sher Khan, Long N. Le, Abhinau K. Venkataramanan +2
The University of Texas at Austin Austin, TX, USA · Samsung Research America Plano, TX, USA
Self-supervised learning (SSL) currently drives state-of-the-art performance in no-reference image quality assessment (NR-IQA). However, standard SSL pipelines uniformly apply synthetic distortions across the entire image field, which can limit their sensitivity to spatially localized and co-occurring degradations encountered in real-world content. In this work, we empirically expose this representational blind spot across existing state-of-the-art encoders, demonstrating their reduced sensitivity to spatially bounded image degradations. To bridge this gap, we introduce Spatial Localized Image Degradation Embeddings for Image Quality Assessment (SLIDE-IQA). SLIDE-IQA employs a dual-branch Vision Transformer framework that injects spatially bounded degradations into a contrastive pretraining objective. To handle the spatial complexity of these degradations, we introduce a Threshold-Bounded Exclusion Mechanism, a representational design choice that resolves structural conflicts arising from spatially localized distortions to ensure the latent space respects both degradation type and spatial scale. Finally, we show that SLIDE-IQA's synthetic-only pretraining significantly improves sensitivity to localized distortions, while achieving competitive performance on NR-IQA benchmarks against existing SSL NR-IQA models.