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.
Figures & tables
Fig. 1: Different enhancement algorithms may produce distinct visual styles, causing algorithm-dependent bias in quality representation. The proposed preference-guided debiasing framework explicitly models enhancement preference and removes its interference, leading to more reliable perceptual quality prediction.
Fig. 2: Overall framework of the proposed method. First, a preference encoder is pretrained with supervised contrastive learning to capture enhancement-preference representations. Second, the learned preference cue is fed into a bias predictor to estimate the bias vector, which is then subtracted from the raw quality feature to obtain a debiased representation for MOS regression.
Fig. 3: Illustration of the proposed preference-guided debiasing method for enhancement image quality assessment. The model first learns a preference embedding from enhanced images to characterize algorithm-dependent enhancement styles, then predicts the corresponding bias term and removes it from the raw quality feature, yielding an algorithm-invariant quality representation for more reliable MOS prediction.
Methods
Type
SRCC
PLCC
KRCC
NIQE [ 10 ]
Handcrafted
0.0200
0.0040
0.0240
BRISQUE [ 9 ]
Handcrafted
0.0302
0.0463
0.0271
ILNIQE [ 16 ]
Handcrafted
0.0956
0.0972
0.0565
IS [ 45 ]
Handcrafted
0.0869
0.0898
0.0483
DBCNN [ 21 ]
Deep-learning
0.4631
0.5131
0.3640
HyperIQA [ 11 ]
Deep-learning
0.6158
0.6702
0.4828
TABLE I: Comparison with state-of-the-art no-reference IQA methods on the RLIE dataset.
Methods
Type
SRCC
PLCC
KRCC
NIQE [ 10 ]
Handcrafted
0.1378
0.1572
0.0550
BRISQUE [ 9 ]
Handcrafted
0.1532
0.1813
0.0593
ILNIQE [ 16 ]
Handcrafted
0.1807
0.1975
0.0610
IS [ 45 ]
Handcrafted
0.1631
0.1760
0.0574
DBCNN [ 21 ]
Deep-learning
0.7934
0.8056
0.6373
HyperIQA [ 11 ]
Deep-learning
0.7419
0.7676
0.5769
TABLE II: Comparison with state-of-the-art no-reference IQA methods on the SQUARE-LOL dataset.
SRCC ↑
PLCC ↑
Methods
Standard
Unseen
Drop ↓
Standard
Unseen
Drop ↓
DBCNN [ 21 ]
0.7934
0.7406
0.0528
0.8056
0.7626
0.0430
HyperIQA [ 11 ]
0.7419
0.6911
0.0508
0.7676
0.7132
0.0544
MUSIQ [ 12 ]
0.7729
0.7424
0.0305
0.7942
0.7644
0.0298
UNIQUE [ 46 ]
0.7766
0.7309
0.0457
0.7899
0.7597
0.0302
TReS [ 23 ]
0.7803
0.7106
0.0697
0.8002
0.7271
0.0731
TABLE III: Cross-algorithm generalization evaluation on the SQUARE-LOL dataset under the algorithm-disjoint protocol. “Standard” denotes the result under the default evaluation protocol, “Unseen” denotes the result on unseen enhancement algorithms, and “Drop” is the performance decrease from Standard to Unseen. The best results are highlighted in bold.
Method
SRCC ↑
PLCC ↑
KRCC ↑
w/o Preference
0.8557
0.8689
0.6802
+ Preference Concat
0.8616
0.8773
0.6834
+ Preference-Guided Debiasing
0.8726
0.8913
0.6904
TABLE IV: Effect of preference-guided debiasing on the SQUARE-LOL dataset. The best results are highlighted in bold.
Method
SRCC ↑
PLCC ↑
KRCC ↑
w/o Preference
0.8557
0.8689
0.6802
Cls-based Preference
0.8701
0.8865
0.6892
SupCon-based Preference (Ours)
0.8726
0.8913
0.6904
TABLE V: Effect of preference representation learning on the SQUARE-LOL dataset. The best results are highlighted in bold.
Training Strategy
SRCC ↑
PLCC ↑
KRCC ↑
Joint Training
0.8510
0.8645
0.6809
Two-Stage w/o Freezing
0.8673
0.8792
0.6864
Two-Stage w/ Freezing (Ours)
0.8726
0.8913
0.6904
TABLE VI: Effect of different training strategies on the SQUARE-LOL dataset. The best results are highlighted in bold.
Sampling Strategy
SRCC ↑
PLCC ↑
KRCC ↑
Random Sampling
0.8527
0.8656
0.6868
Algorithm-Balanced Sampling
0.8578
0.8690
0.6863
Content-Controlled Sampling (Ours)
0.8726
0.8913
0.6904
TABLE VII: Effect of different sampling strategies for preference learning on the SQUARE-LOL dataset. The best results are highlighted in bold.
Fig. 4: Qualitative comparison on images enhanced by different algorithms. Images in each row share the same source content but exhibit different enhancement styles. Although their MOS values are relatively similar, the model without preference guidance produces inconsistent predictions, while the proposed method yields predictions that are more stable and closer to the ground-truth MOS. Red numbers indicate relatively large prediction errors.