Deep Opinion-Unaware Blind Image Quality Assessment by Learning and Adapting from Multiple Annotators

1 Introduction

Most reliable methods for automatic image quality assessment (IQA) are "full-reference", relying on the availability of the reference pristine image [Wang et al., 2004]. Nonetheless, in many practical applications, the reference image is not accessible or even not in existence, necessitating the use of no-reference or blind IQA (BIQA) models [Mittal et al., 2012]. The working mechanism of BIQA is close to our vision system since humans are able to assess the perceptual quality of distorted images without referring to their pristine counterparts. Presumably, our visual systems have, through evolutionary and developmental processes, learned to preferentially encode the prior knowledge of the "naturalness" of visual images - those may arise from the physical interactions of light, surfaces, and optics. Early attempts at BIQA models are mainly based on natural scene statistics (NSS), with the underlying assumption that natural images are innate with a number of quality-aware consistent statistical regularities. Spatially normalized coefficients and codebook-based representations are two representative pipelines that have achieved impressive performance on generic distortion types, e.g. blur, noise, and JPEG compression [Mittal et al., 2012].

∗Corresponding author.

When generalizing to in-the-wild distortions, statistical regularities may not accurately model these distortions, resulting in poor performance. Recently, BIQA methods based on deep neural networks (DNNs) have shown remarkable performance, surpassing traditional NSS-based methods [Zhang et al., 2018]. This success is attributed to the ability of DNNs to automatically learn informative feature representations for quality prediction. Training DNNs with millions of parameters typically requires extensive human quality annotations, such as mean opinion scores (MOSs). The acquisition of large-scale human-labeled datasets is both laborious and costly. For instance, the KonIQ-10k dataset [Hosu et al., 2020] comprises over 10,000 images annotated by nearly 1,500 subjects, resulting in 1.2 million human ratings. To mitigate the reliance on human annotations, some studies explore the use of pseudo-labeled datasets to train DNN-based BIQA [Ye et al., 2014; Ma et al., 2017]. These approaches effectively mitigate the reliance on human annotations.

In this study, we further explore the use of synthetically distorted images with pseudo labels [Ma et al., 2019] for BIQA by proposing a Deep opinion-Unaware BIQA model that learns and adapts from Multiple Annotators, termed DUBMA. Specifically, we begin by generating a large-scale set of distorted image pairs, for which a collection of fullreference IQA (FR-IQA) models provides relative quality rankings. We then employ a learn-to-rank (L2R) approach [Wang and Ma, 2021] to learn from these FR-IQA models. During the training of DUBMA, each annotator is assigned a distinct reliability score, which is implicitly estimated via maximum likelihood estimation (MLE), thereby enabling robust label fusion. Previous studies, such as BLISS [Ye et al., 2014] and Wu et al.[Wu et al., 2020], have investigated the use of full-reference (FR)-IQA methods to generate quality scores for unlabeled images. BLISS employs unsupervised rank aggregation to combine scores from multiple FR-IQA models, while Wu et al.[Wu et al., 2020] scale predictions from various FR-IQA models to a common range to produce pseudo-MOS. However, these approaches have a significant limitation: they do not consider the reliability of individual FR-IQA models, which can introduce noise into the aggregated scores due to varying prediction accuracies across different distortion types.

Due to the inherent discrepancies between synthetic and real-world distortions, a domain shift may arise, causing BIQA models trained solely on synthetic distortions to struggle in accurately assessing realistic camera distortions [Zhang et al., 2021]. To overcome this challenge, we adopt unsupervised domain adaptation (UDA) methods [Borgwardt et al., 2006], which have demonstrated great potential for enhancing model performance on unlabeled target domains by effectively transferring knowledge from source domains. At the heart of adapting from synthetic distortions (as the source) to realistic distortions (as the target) is the usage of optimal transport (OT) distances to quantify the distance between distortions, more precisely, the distance between feature-label pairs across domains [Flamary et al., 2017]. This property is particularly applicable to our scenario, since it is difficult to ensure that the quality range of the synthetic images completely covers realistic ones.

In summary, this work makes two primary contributions:

• We propose DUBMA, an opinion-unaware BIQA method trained on synthetic data pseudo-labeled by multiple FR-IQA models. The framework incorporates a L2R strategy, assigning distinct reliability scores to each annotator to enable joint learning from multiple noisy FR-IQA models. The resulting model achieves robust performance on synthetic distortions.
• We integrate an outlier-robust UDA approach to adapt BIQA models trained on synthetic data for realistic distortions. Using unbalanced OT distances, we quantify discrepancies between features and labels across domains. The adapted models demonstrate strong generalization and effective handling of real-world distortions.
• We conduct extensive experiments across multiple human-rated IQA datasets and perform detailed ablation studies to validate the effectiveness and robustness of our opinion-unaware BIQA learner in challenging realworld scenarios.

2 RelatedWork

2.1 Opinion-Unaware BIQA

Early opinion-unaware BIQA methods primarily rely on quality-aware natural scene statistics (NSS) features. Mittal et al. [Mittal et al., 2011] introduced TMIQ, which assesses image quality by comparing the similarity of NSS features between distorted and pristine images. Later, Mittal et al. [Mittal et al., 2012] developed NIQE, which models the distribution of natural NSS features using a multivariate Gaussian (MVG) model. Similarly, Zhang et al. [Zhang et al., 2015] extended the quality-aware NSS features to propose ILNIQE. With the advent of DNNs, some studies have explored learning from pseudo-labels to train opinion-unaware BIQA models. Ye et al. [Ye et al., 2014] introduced BLISS, which utilizes synthetic scores from full-reference IQA methods to train BIQA models. Liu et al. [Liu et al., 2017] proposed RankIQA, where distortion intensities are used to determine the ranking order of image pairs. However, these methods, mainly designed for synthetic distortions, struggle with generalizing to realistic ones. In contrast, Zhu et al. [Zhu et al., 2021] employed the trained discriminator of a Wasserstein GAN for quality assessment of realistic distortions. While they suggest that the discriminator may capture real-world image distributions, the performance remains suboptimal.

2.2 Deep Unsupervised Domain Adaptation

Popular UDA methods are primarily classified into statistical moment matching [Borgwardt et al., 2006; Sun et al., 2016; Damodaran et al., 2018], domain style transfer, self-training, and feature-level adversarial learning [Ganin et al., 2016; Shen et al., 2018]. The testing benchmarks for deep UDA mainly focus on the image classification task. In the IQA/VQA setting, Chen et al. [Chen et al., 2021a] developed the first Maximum Mean Discrepancy (MMD)-based domain adaptation for the quality assessment of screen content images (SCIs). Chen et al. [Chen et al., 2021b] proposed a curriculum-style unsupervised domain adaptation to handle the cross-domain problem for no-reference video quality assessment. Recently, optimal transport (OT) distance, particularly Wasserstein distance [Flamary et al., 2017], has gained attention for UDA. Flamary et al. [Flamary et al., 2017] proposed a regularized OT method for aligning representations between source and target domains. However, OT treats all samples, including outliers, equally due to marginal constraints, which can lead to outliers disproportionately affecting the total mass. To address this, Benamou et al. [Benamou, 2003] introduced a penalization term to relax these constraints, resulting in unbalanced OT for handling unbalanced data. Furthermore, Fatras et al. [Fatras et al., 2021] showed that unbalanced OT reduces bias caused by minibatches.

3 ProposedMethod

Figure 1 illustrates the overall framework of DUBMA. Our learning paradigm assumes the existence of a training set consisting of ns synthetic image pairs Ds = {(x(i) s , y(i) s )}ns i=1 and a test set Dt = {x(j) t }nt j=1 of nt images. Given an instance xs ∈ Ds, its true perceptual quality is denoted by q(xs). We leverage M objective IQA models {qm}Mm=1 to compute the quality prediction of q(xs), which are collectively denoted by {qm(xs)}Mm=1 . Since different objective IQA models may exhibit varying degrees of non-linearity and scale, we adopt the L2R strategy, where pseudo labels are organized in binary form to avoid being affected by nonlinearity and scale, thereby mitigating these effects [Zhang et al., 2021]. Specifically, given an image pair (xs, ys) ∈ Ds, we assign a binary label rm for this pair, where rm = 1 if qm(xs) > qm(ys) and rm = 0 otherwise. Therefore, the pseudo-labeled training set is in the form of Ds = {(x(i) s , y(i) s ), r(i) 1 , · · · , r(i) M }ns i=1 . We aim at training a DNNbased BIQA method on Ds to reliably estimate the perceptual quality of the sample in Dt. Such a function includes two mappings: a differentiable embedding function g, which maps the input into the latent embedding, and the quality predictor h, which maps the embedding to the quality score.

3.1 DUMBA for Synthetic Distortions

Since the quality estimate of each IQA annotator is noisy and not equally good (or bad) at labeling the input, we introduce a set of the reliability parameters α = {αm}Mm=1 to explicitly model the hit rate and correct reject rate of each annotator, which are denoted as:

where r = 1 if q(xs) ≥ q(ys) and r = 0 otherwise. Under the hypothesis of Thurstone's Case V model [Zhang et al., 2021], we assume q(xs) ∼ N (h(g(xs)), 1). The perceptual difference between a stimuli pair (xs, ys) takes the form q(xs) - q(ys) ∼ N (h(g(xs)) -h(g(ys)), 2). From that, the strengths (probability) of preferring xs over ys perceptually in an image pair (xs, ys) can be computed from the Gaussian cumulative distribution function (CDF) Φ(·), which has a closed form as:

During training, we sample a mini-batch Bs from Ds in each iteration and minimize the negative logarithm maximum likelihood function [Ma et al., 2019] to obtain the optimal model parameters along with the reliability variables α = {αm}Mm=1 , which is defined as:

where

The resulting DUBMA model is capable of performing reliable quality assessment for synthetic distortions.

3.2 DUMBA for Realistic Distortions

The distribution shift between synthetic and realistic distortions may lead to underperformance of DUBMA, which is trained on synthetic data, when applied to realistic distortions. To address this issue, we incorporate UDA to align the distribution shift (see Figure 1).

Optimal Transport. The OT problem is concerned with finding the minimum mass of transport required to transform a source probability distribution Ps to a target probability distribution Pt. Specifically, we assume that δ(x) denotes the Dirac function at location x, p(i) s and p(j) t respectively represent the probability masses associated with the i-th source element x(i) s and j-th target element x(j) t , belonging to the probability simplex, i.e., ∑ns i=1 p(i) s = ∑nt j=1 p(j) t = 1. The empirical marginals ps and pt are written as:

where ns and nt represent the number of source and target samples. The Kantorovitch's relaxation of the OT problem in discrete setting is formulated to seek the optimal transport plans or couplings γ ⋆ as:

where Π(ps, pt) = {γ ≥ 0|γ1ns = ps, γ T1nt = pt} (1d is a d-dimensional vector of ones) is a set of all joint distributions (couplings), ⟨·, ·⟩F denotes the Frobenius dot product, C ≥ 0 is the ns × nt cost function matrix, whose term Cij = c(x(i) s , x(j) t ) measures the pairwise cost to move a probability mass from x(i) s to x(j) t . The minimal cost can be used to quantify the distance between two distributions and is referred to as the Kantorovich-Rubinstein-Wasserstein distance, i.e.,

Unbalanced Joint OT. The fundamental idea behind joint distribution alignment is to address the variations in both feature space and conditional label space [Courty et al., 2017; Damodaran et al., 2018]. However, since true quality labels are not available in either the source or target domains, we substitute them with quality predictions. To be more specific, assuming we have access to the source and target datasets Ds and Dt respectively, we redefine the joint empirical marginals ps and pt as:

where g and h represent the feature extractor and quality estimator, respectively. The cost function c in Eq (7) needs to measure the distance between the features and the discrepancy between predictions in combination, which is defined as:

where the second term measures the fidelity between p(xs, ys) and r = 0.5, and β is the trade-off to balance two terms.

In practical OT optimization, each sample is weighted equally due to the marginal constraints. This means that p(i) s = 1/ns for all source samples and p(j) t = 1/nt for all target samples. However, OT is sensitive to outliers due to the distribution's geometry. In our framework, it is inevitable to generate synthetic data with distributions that are far away from realistic data. To overcome this issue, we can introduce a "soft" penalty term that relaxes the "hard" marginal constraints. This assigns smaller weights to outliers compared to in-distribution samples, effectively ignoring them. This formulation is known as unbalanced OT, which is defined as:

where ˆ ps and ˆ pt are the plan's marginals of unbalanced OT, λ is the tradeoff that represents the strength of penalization;

KL is the Kullback-Leibler (KL) divergence. It should be noted that the marginals of γ are no longer equal to (ps pt) in general. In DNN training, OT is computed over minibatches, with results averaged to approximate the full dataset's OT. However, minibatch sampling can introduce deviations between the true OT loss and its estimate [Fatras et al., 2021]. Unbalanced OT addresses this by relaxing marginal constraints, improving robustness to minibatch effects and scaling well with large datasets.

Optimization of Unbalanced Joint OT. Directly optimizing the unbalanced OT is nontrivial and ill-posed. We will reformulate the optimization problem (10) as a non-negative penalized linear regression problem [Chapel et al., 2021]. Let t = vec(γ), c = vec(C) and p⊤ = [p⊤s , p⊤t ]. Problem (10) can be re-written as:

where H⊤ = [H⊤ r ,H⊤ c ] concatenates the sums of the rows and columns of γ, respectively; KL(Ht, p) denotes the data fitting term. This formulation is well-known in inverse problems and non-negative matrix factorization [Lee and Seung, 2000]. Here, we leverage the majorization-minimization (MM) algorithm to solve the problem in the form (11), which leads to efficient multiplicative updates for the KL penalty [Chapel et al., 2021].

Optimization Objective over Minibatches. The proposed BIQA model is trained in an end-to-end manner using a variant of the stochastic gradient descent method. Specifically, we sample mini-batches Bs and Bt from Ds and Dt, respectively, in each iteration. Then the empirical loss function on Bs and Bt is defined as:

ℓq is defined in Eq. (3), p(i) s = 1/ns for all i, p(j) t = 1/nt for all j, and η trades off the two terms. In Eq. (12), there are two sets of parameters to optimize: the OT couplings γ and the model parameters {h, g,α}. During training, we use the alternative minimization approach [Damodaran et al., 2018] to update these parameters. Firstly, we solve the optimal transport plan γ ⋆ with fixed model parameters {h, g} using nonnegative penalized linear regression. Then, we optimize the model parameters {h, g, α} using stochastic gradient descent with the obtained γ ⋆ .

4 Experiments and Results

4.1 DUMBA for Synthetic Distortions

Datasets. We build the pseudo-labeled dataset Ds based on the Waterloo Exploration Database [Ma et al., 2016]. We take Counts these images as references and simulate twenty-five distortion types 1 , each at five levels. We employ ten state-of-theart (SOTA) FR-IQA models for pseudo labeling i.e., SSIM [Wang et al., 2004], MS-SSIM [Wang et al., 2003], SR-SIM [Zhang and Li, 2012], FSIM [Zhang et al., 2011], VSI [Zhang et al., 2014], VIF [Sheikh and Bovik, 2006], GMSD [Xue et al., 2013b], MDSI [Nafchi et al., 2016], NLPD [Hadizadeh et al., 2017], and A-DISTS [Ding et al., 2021]. We select these models based on their performance evaluated by D-test, Ltest, and P-test on the generated dataset [Ma et al., 2016]. We generate over 90k image pairs and randomly select an appropriate number for training. Figure 2 illustrates the consistency of the IQA annotations across all pairs, counting the number of models that selected the first image in each pair as the higher quality one. Overall, the ten IQA annotators show high agreement, with complete consistency on roughly 80% of the image pairs. We evaluate the DUBMA performance on three commonly used subject-rated image sets - LIVE [Sheikh et al., 2006], CSIQ [Larson and Chandler, 2010], and KADID10k [Lin et al., 2019].

Figure 2: Histogram of the number of IQA models voting a "1".

Details of Training and Testing. We use ResNet-18 pretrained on ImageNet as the backbone g for feature extraction, followed by two fully connected layers with ReLU activation as the quality predictor h. The model is trained on Ds using the Adam optimizer, minimizing the loss in Eq. (3). The learning rates for g, h and α are set to 10-4 and 10-3 , respectively, with a decay factor of 2 every two epochs. The batch size is 64, and training lasts for seven epochs. The best model is selected based on the average entropy of 5,000 randomly sampled image pairs from KADID-10k, with the model showing the least entropy chosen for performance comparison. We utilize the Spearman rank correlation coefficient (SRCC) and the Pearson linear correlation coefficient (PLCC) to quantitatively assess the model's performance. In addition to this, we employ the group MAximum Differentiation (gMAD) competition [Zhang et al., 2021] to qualitatively

1 These include Gaussian blur, lens blur, motion blur, color diffusion, color shift, color quantization, color saturation 1/2, JPEG2000 compression, JPEG compression, white noise, whiter noise in color component, impulse noise, multiplicative noise, denoise, brighten, darken, mean shift, jitter, non-eccentricity patch, pixelate, quantization, color block, high sharpen, and contrast change.

Figure 3: Representative gMAD pairs between DUBMA and MDFS [Ni et al., 2024] on synthetic distortions. (a) Fixing DUBMA at the low-quality level. (b) Fixing DUBMA at the high-quality level. (c) Fixing the MDFS at the low-quality level. (d) Fixing the MDFS at the high-quality level.

evaluate method's generalizability.

Quantitative Results. We compare the proposed method with eight knowledge-driven BIQA methods - QAC [Xue et al., 2013a], PIQE [Venkatanath et al., 2015], LPSI [Wu et al., 2015], NIQE [Mittal et al., 2012], ILNIQE [Zhang et al., 2015], SISBLIM [Gu et al., 2014], SNP-NIQE [Liu et al., 2019], NPQI [Liu et al., 2020], and eight data-driven BIQA models - RankIQA [Liu et al., 2017], dipIQ [Ma et al., 2017], CaHDC [Wu et al., 2020], EONSS [Wang et al., 2019], Ma19 [Ma et al., 2019] Zhu21 [Zhu et al., 2021], ContentSep [Babu et al., 2023], MDFS [Ni et al., 2024]. All of these models are opinion-unaware. We also include three DNN-based full-reference IQA methods - PieAPP [Prashnani et al., 2018], LPIPS [Zhang et al., 2018], and DISTS [Ding et al., 2022], trained on human-labeled datasets as reference. The implementations of all these methods are obtained from the respective authors. Table 1 compares our method with other BIQA models. We observe DUBMA achieves significant improvements on KADID-10k [Lin et al., 2019], demonstrating the effectiveness of our learning framework. Unlike CaHDC [Wu et al., 2020] and EONSS [Wang et al., 2019], which use equally weighted pseudo-MOSs from FRIQA models, DUBMA outperforms these methods by assigning varying reliability levels to each annotator, reducing the impact of noisy labels. Additionally, DUBMA, trained on a larger dataset with more accurate FR-IQA annotators, shows better generalization on LIVE and KADID-10k compared to Ma19 [Ma et al., 2019], which uses a smaller model. Remarkably, our method surpasses PieAPP [Prashnani et al., 2018] and LPIPS [Zhang et al., 2018] without relying on human data and matches the performance of DISTS [Ding et al., 2022] on KADID-10k, proving the effectiveness of learning from pseudo-labels.

gMAD Results. We conduct a gMAD competition game between DUBMA and MDFS [Ni et al., 2024], the secondbest-performing BIQA model on the synthetic dataset. The competition takes place on the Wang and Ma [Wang and Ma, 2021] dataset, which includes 100,000 images generated from 1,000 different scenes. Fig 3 illustrates the gMAD pairs between DUBMA and MDFS. In (a) and (b), where DUBMA and MDFS act as the defender and attacker, respectively, the top images have similar quality compared to the corresponding bottom images, indicating that DUBMA successfully resists the attack fromMDFS. When the roles are reversed (see (c) and (d)), DUBMA consistently identifies the failure cases of MDFS, with pairs of images showing substantial differences in quality according to human perception. These findings suggest that DUBMA generalizes better than existing opinion-unaware BIQA models on synthetic distortions.

4.2 DUMBA for Realistic Distortions

Datasets. We also leverage the Waterloo Exploration Database [Ma et al., 2016], and simulate five distortion types

2 at five levels to construct Ds. The pseudo labels are annotated by the aforementioned ten FR-IQA models. We generate about 25k image pairs for training. We use KonIQ-10k [Hosu et al., 2020] and SPAQ [Fang et al., 2020] as the target datasets. Both of them include more than 10, 000 authentically distorted images.

Details of Training and Testing. We use the same network architecture as the experiment on synthetic distortions. We train our model on the synthetic dataset Ds and adapt to the realistic dataset Dt by minimizing the loss function of Eq. (12). We rely on the POT package [Flamary et al., 2017] to compute the exact OT solver and the Geomloss package for the unbalanced Sinkhorn divergence. The entropy of the probability output is employed as the criterion for validation. We also adopt the SRCC and PLCC for accuracy evaluation and gMAD competition to assess the generalization capacity. Quantitative Results. We first compare the performance of the proposed method against the aforementioned sixteen BIQA models (see Section 4.1). Table 2 lists the SRCC and PLCC results between model predictions and MOSs on KonIQ-10k [Hosu et al., 2020] and SPAQ [Fang et al., 2020], respectively. We also include the performance on LIVE as a reference to measure the performance on synthetic distortions. It is easily observed that the BIQA methods initially tailored for synthetic distortions work well on LIVE but perform poorly on two realistic datasets with diverse content and complex distortions, especially the catastrophic performance Correlation Correlation 0.4 0.6 0.8 Correlation 0.6 0.7 0.8 0.9 Correlation of QAC. One possible reason is that the NSS is inadequate to model realistic distortions, characterized by complex image content and distortion types. Our proposed DUBMA achieves competitive performance on these three datasets, which confirms the effectiveness of our proposed method.

2 These include Gaussian blur, Gaussian noise, motion blur, JPEG compression, and JPEG2000 compression.

(b) Number of distortion levels (c) Number of annotators 2 4 6 8 10 Number of annotators 4 5 Number of distortion levels that the performance is improved with the increase of annotators, and it is best when the number of annotators is 10. Thus, we could leverage more reliable annotators to boost model performance. We finally verify the effectiveness of joint adaptation of both the marginal distribution and conditional distribution between domains [Courty et al., 2017; Damodaran et al., 2018]. Table 4 reports the comparison results obtained using the joint or non-joint cost function. We observe an obvious performance gain achieved by jointly aligning feature and conditional label space. This gain might indicate that only adapting the marginal distributions is not enough for the transfer learning of the proposed model.

(a) Number of distortion types 1 5 13 25 Number of distortion types 0.2 2 3 0.5 We then compare unbalanced OT against the six frequently used SOTA UDA methods - CORAL [Sun et al., 2016], DANN [Ganin et al., 2016], MMD [Borgwardt et al., 2006], WDGRL [Shen et al., 2018], DeepJDOT [Damodaran et al., 2018] and SWD [Lee et al., 2019]. CORAL, DANN, MMD, and WDGRL perform domain alignment on the feature space, while DeepJDOT and SWD minimize the discrepancy of joint deep feature and label domains. Table 3 reports the comparison results, where source only means that the BIQA model is only trained on Ds directly. We find UDA methods bring obvious performance gains compared to source only, except in the case of CORAL on SPAQ. Besides, joint alignment of both feature and label space is slightly superior to only matching feature space. Moreover, the proposed method outperforms all competing models by a significant margin.

gMAD Results. We qualitatively evaluate DUBMA and second-best MDFS [Ni et al., 2024] using the gMAD competition. The competition performs the Wang et al. [Wang et al., 2021] dataset, which includes 100,000 authentic images collected from the Internet. Figure 4 shows representative gMAD pairs between the two models. Initially, we allow MDFS to identify the failure cases of DUBMA, and we find that MDFS fails to detect these cases. In contrast, when we reverse the roles, DUBMA successfully identifies the failure cases of MDFS (see (c) and (d)). These results indicate that DUBMA outperforms MDFS for in-the-wild BIQA tasks.

Ablation Studies. We first probe the effect of the number of distortion types, and Figure 5 (a) shows the results. It could be observed that even synthesizing only one type of distortion to construct Ds, the BIQA model can achieve impressive performance. With the increase of distortion types, the performance on KonIQ-10k and SPAQ is influenced slightly. We argue that UDA benefits BIQA models in learning distortion-unaware features. We then test the influence of distortion levels highly related to the image quality ranges. Figure 5 (b) shows the performance changes versus distortion levels. As we can see, the model performance is improved with more distortion levels, and five distortion levels achieve the relatively best performance. We further study the impact of the number of annotators. The results can be found in Figure 5 (c). It could be observed

5 Conclusion

We propose DUBMA in this paper that is trained on scalable pseudo-labeled synthetic pairs to handle both synthetic and real distortions. The approach uses supervised signals annotated by multiple existing FR-IQA models, eliminating the need for human ratings. To address the domain shift between the synthetic source data and the realistic target test data, we use robust joint UDA to align the two distributions. This allows us to train BIQA models on large-scale datasets that can be easily collected and scaled up without human input. Our extensive experiments demonstrate that our optimized model performs favorably against current opinion-unaware BIQA models. However, while the model's performance on authentic databases is promising, it has not yet achieved the performance level of models trained on human-rated datasets. We expect to improve the model performance by incorporating additional reliable IQA annotators and leveraging more efficient UDA solutions.

Contributions

Zhihua Wang and Xuelin Liu are co-first authors.

Acknowledgements

This work was supported in part by the National Natural Science Foundation of China (62301323, 62301621,

62461028), in part by the Shenzhen Science and Technology Program (20231121172359002, JCYJ20241202125904007), in part by the Guangdong Basic and Applied Basic Research Foundation (2024A1515011164, 2025A1515011398), and in part by the China Postdoctoral Science Foundation (2024T170364).