Thank you for your affirmation and the valuable comments that help us improve our work. The following is a careful response and explanation about the weaknesses and questions.

For Experimental Designs Or Analyses

We sincerely appreciate your recognition of our experimental design and analysis, as well as this valuable suggestion.

In fact, we have already applied the proposed method to train an IQA model using a combination of existing annotated IQA datasets and additional unlabeled images under standard experimental settings. Specifically, as shown in Appendix Table 10, we select 80% from BID,70% from KonIQ-10k, 10% from KonIQ10k,and the remaining 20% from two datasets as labeled, unlabeled, validation, and testsets, respectively. The results clearly demonstrate the superiority of our approach.

To further validate the effectiveness of our method, as suggested by the reviewer, we split KonIQ-10k into labeled training and test sets (8:2 ratio), trained the model on the labeled KonIQ-10k training set and the unlabeled LIVE-C dataset, and evaluated it on the remaining 20% KonIQ-10k test set. The results are as follows:

From the above results, we can observe that, after training with the extra unlabeled dataset LIVE-C, the performance of the model is improved on the test set of KonIQ-10k compared to the counterpart which is trained only on training set of KonIQ-10k. These results further validate the effectiveness of our proposed method.

In the camera-ready version, we will incorporate these additional experimental findings to further strengthen our work.

For Other Strengths And Weaknesses

Thanks for this valuable comment. We acknowledge that resizing could theoretically alter distortion characteristics and affect quality assessment accuracy. In fact, the resize-and-crop strategy in our model is actually another choice that is widely used in many works, which can balance computational efficiency and practical applicability. Below, we provide a detailed justification for this design choice in our model from three perspectives:

1. Preservation of Global Context: Since the required final cropped size during training is 224×224, resizing images to 256×256 before cropping in our method ensures that the model captures relatively global regions of the image while retaining fine-grained distortion patterns. This strategy has been widely adopted in classic IQA works such as DB-CNN, NIMA, and SSL-IQA, and our approach follows this well-established practice.
2. Trade-off Between Performance and Efficiency: We agree with the reviewer that cropping multiple patches from original-resolution images (without resizing) is theoretically preferable. However, this requires extracting multiple patches per image, significantly increasing training overhead and computational complexity. This is because, for large original images, a small number of cropped patches may fail to cover sufficient global regions of the image. Therefore, existing models adopting this strategy (e.g., LIQE and HyperIQA) typically address this by cropping a large number of patches, which further escalates computational costs.
3. Empirical Validation: Following the experimental setup in Table 1, we conducted additional tests by randomly cropping 10 patches per original-resolution image (denoted as Strategy 2) and compared its performance with our default resize-and-crop approach (Strategy 1). The results show that Strategy 2 only improves PLCC and SRCC by less than 1%, but at the cost of 4× longer runtime. The cost-performance ratio is unfavorable.

We will emphasize in camera-ready version that our resizing-and-cropping strategy balances efficiency and performance, ensuring scalability for real-world deployment.

For Questions

Thank you for your careful review and valuable feedback. We are sorry for this oversight and will correct it in the camera-ready version by bolding the SRCC result of HyperIQA in the table.

Notably, this oversight does not impact Appendix E.5's analysis, as Table 10 primarily compares our method with Semi-IQA, where ours shows superior performance. While HyperIQA achieves marginally better results on BID through computationally intensive supervised training, it underperforms significantly on KonIQ-10k compared to our method, further validating our approach's effectiveness.

In camera-ready version, we will conduct a more thorough review to eliminate any remaining typos and ensure the rigor and high standard of our research.

Thanks for your affirmation and the valuable comments that help us improve our work. The following is a careful response and explanation about the weaknesses and questions.

For Weakness 1

We sincerely appreciate the constructive feedback. We would like to clarify that Sections 3.3.1 to 3.3.6 (pages 4–6) of the manuscript have already provided a comprehensive and step-by-step explanation of the algorithm’s implementation.

To further enhance clarity and ensure readers can better understand the implementation details, in the final version, we will include more illustrative examples or pseudocode snippets within the main text to complement the existing Algorithm 1 in Appendix C, which can further improve the readability and transparency of the method while maintaining the rigor of the technical content.

For Weakness 2

Thank you for the insightful suggestion. In the final version, we will add a concise discussion in Section 3.3.7 for a more concise exposition of the intuitive understanding of these theoretical linkages, including the following two key insights:

On the one hand, the manifold assumption serves as the theoretical foundation for label propagation, positing that samples with similar features (i.e., neighboring nodes in the graph) should share similar labels. This fundamental assumption enables the label propagation process, where labels from annotated samples are effectively disseminated to unlabeled ones through the constructed nearest neighbor graph (Eq.4), thereby facilitating pseudo-label learning. On the other hand, for the optimization perspective, the iterative propagation (Eq. 5) implicitly solves the regularization problem (Eqs. 13-14), where the first term enforces smoothness (labels vary slowly over the constructed nearest neighbor graph), and the second term fits the initial labels. The hyperparameter $\mu$ in Eq.14 balances the above two objectives.

For Questions

Thank you for the valuable comments that help us improve our work. In fact, both of the proposed model and label conversion strategy exhibit strong robustness across diverse datasets with varying MOS distributions, which have been comprehensively demonstrated in our paper. Specifically:

On the one hand, we have conducted experiments on multiple datasets (KonIQ-10K, SPAQ, LIVE-C, and NNID) with significantly different MOS ranges and distribution characteristics (e.g., KonIQ-10K: [1,5], SPAQ: [0,100]). The consistent performance improvement (Tables 1–3) across these datasets demonstrates the generalization ability of our label conversion strategy.

Additionally, in Appendix E.5 and E.6, we explicitly tested scenarios where unlabeled training data came from different distortion types (authentic vs. synthetic) and sources (e.g., training with labeled dataset BID and unlabeled datasets KonIQ-10K in Appendix E.5, and training with labeled dataset KonIQ-10K and unlabeled datasets SPAQ/KADID-10K in Appendix E.6). Results in Tables 10 & 11 show that our method remains effective even when MOS distributions differ, as the entropy minimization inherently adapts to the input label range (Eq. 1 normalizes MOS to [1,100]) and preserves relative quality relationships. The confidence-weighted training (Eq. 8–9) further mitigates potential outliers from label conversion.

To further validate the robustness of our proposed strategy, we conducted additional experiments: we split KonIQ-10k into labeled training and test sets (8:2 ratio), trained the model on the labeled KonIQ-10k training set and the unlabeled LIVE-C dataset, and evaluated it on the remaining 20% KonIQ-10k test set. The results are as follows:

The above results show that our method is robust to different datasets.

On the other hand, Although the label conversion process in Eq. (2) cannot be directly applied to predict MOS distributions (since our method operates under the reasonable assumption that the labeled samples have sufficiently high confidence levels (Lines 190-192)), the well-trained CPL-IQA model can robustly predict label distributions. For example, using the model trained as described above, we randomly selected 10 samples each from KonIQ-10k and LIVE-C, and computed the average JS divergence and Wasserstein distance between the predicted MOS distributions and the variance-simulated normal distributions (approximating ground truth). The results are as follows:

The results demonstrate that both the JS divergence and Wasserstein distance remain below 0.1, confirming the accuracy of the predicted MOS distributions and validating the robustness of our method across diverse datasets and MOS distributions.

In final revision, we will further highlight the robustness.

We appreciate the valuable comments and will improve in final version.

For Claims & Evidence

This is a misunderstanding. We clarify that:

1. The purpose of Fig. 1 is to demonstrate the limitations of some existing semi-supervised BIQA methods that require full score distributions (Left Lines 61-68). As shown, images with similar MOS values can have divergent score distributions, making direct MOS-to-distribution conversion unreliable. Unfortunately, many datasets lack distribution information, limiting these semi-supervised methods' applicability.

2. While some datasets provide variance information, simulating a normal distribution solely based on variance is often inaccurate. Moreover, in real-world scenarios, variance information is frequently unavailable. These limitations restrict the broad applicability of these semi-supervised methods. In addition, as shown in Left Line 95, it is confirmed that models trained with vectorized distribution labels outperform those using scalar MOS labels. Therefore, the two conflict findings motivated us to develop a more flexible solution that converts MOS labels into vector labels for training without requiring distribution and variance.

3. Therefore, the vectorized conversion process in our method does not depend on variance information. To achieve this, we propse the entropy minimization (Eq.2) under the high-confidence MOS assumption (Left Lines 182-201), leading both better performance than scalar MOS training and broader applicability than distribution-dependent methods.

Thus, Fig. 1 and our method are logically consistent. We will clarify this point more explicitly in final version to prevent any potential misunderstanding.

For Weakness 1

We acknowledge that label conversion is relatively common in IQA. However, existing methods are designed specifically for either fully supervised or LLM-based BIQA, including:

1. [1] discretizes the score range into five equal intervals, converting continuous scores into one-hot encoded rating levels.
2. [2] converts MOS labels into vector representations by assuming each MOS score follows a normal distribution, which is simulated using a predetermined fixed variance.
3. [3] extends [2] by incorporating sample-specific variance values.

We conduct experiments comparing our method with: (1) the one-hot conversion in [1], and (2) the normal distribution (N-D) simulation used in [2-3], using the same setting as Tab. 1.

Evidently, existing methods are unsuitable for semi-supervised BIQA scenarios due tofailing reliable pseudo-label confidence assessment. In contrast, our method innovatively addresses this gap.

[1] Q-align. ICML'24

[2] BIQA with Probabilistic Representation. ICIP'18

[3] A fuzzy neural network for opinion score distribution prediction for BIQA. TCSVT'24

For Weakness 2

While our method shows marginal improvement over SSLIQA in Tab 1, it demonstrates significantly superior performance in cross-dataset experiments on LIVE-C in Table 2, highlighting its exceptional generalization capability. Moreover, as detailed in Line 715, SSLIQA requires multi-branch network training, whereas our approach maintains compatibility with any single-branch backbone architecture, offering reduced parameters and simpler structure. These advantages substantiate the significance of our method.

For Weakness 3

Thanks for your suggestion. In fact, the setting $B_L=8, B_U=56$ in Stage 2 is to match Stage 1's batch size $64$, i.e., $B_L+B_U=64$. Moreover, the significantly larger size of $B_U$ compared to $B_L$ reflects the practical scenario where labeled samples are substantially fewer than unlabeled samples.

To investigate the impact of ratio $B_L / B_U$, under the settings of Tab. 1, we obtained the following results:

The results show that while an excessively small $B_L$ may degrade performance; the model exhibits relative insensitivity to variations in the ratio of $B_u / B_L$.

For Weakness 4

This is a misunderstanding. The two-step training in our method does not put a burden on the implementation.

As stated in Sec. 3.2, our framework employs a two-stage approach: Stage 1 utilizes only labeled data for training, introducing no additional computational overhead; Stage 2 consists of two steps, with the pseudo-labeling process requiring merely 2.1% of the total training time (see Response to Reviewer R7vR, Question 2).

Moreover, training on SPAQ, our method requires only 16.69 minutes (1 epoch in Stage 1 + 1 epoch in Stage 2) on a single 2080Ti GPU, versus SSLIQA's 18.31 minutes for 2 epochs. This efficiency gain stems from our single-branch design versus SSLIQA's dual-branch architecture, demonstrating our superior efficiency.

Thanks for the affirmation and valuable suggestions. We have addressed each point below in detail.

For Essential References Not Discussed

In fact, our method is not directly comparable with [1], [2], or [3], as they follow fundamentally different paradigms. Specifically:

• Ours: A semi-supervised BIQA method that leverages confidence-quantifiable pseudo-label learning to utilize unlabeled data.
• [1]: A fully supervised LLM-based approach trained on multiple joint datasets, requiring massive computational resources.
• [2]: A fully supervised method that adapts large-scale pretrained models with minimal trainable parameters.
• [3]: A fully supervised approach using multiple comparison transformers and score pivots to construct an embedding space.

Thus, our method is entirely incomparable with [1] and, strictly speaking, also with [2] and [3]. Nevertheless, for a fair comparison under as equal training conditions and model sizes as possible, we conducted experiments following the settings in Table 2 of our paper:

The results confirm our method's superiority. In final version, we will cite [1], [2], [3] and expand the experimental section to include more detailed discussions and comparisons with them, further validating the effectiveness of our approach.

For Question 1

In fact, as mentioned in Left Lines 423-427, Figure 4 primarily serves to demonstrate that the distribution of predicted pseudo-labels closely aligns with that of the ground truth, thereby validating the effectiveness of our Label Optimizing module.

For quantitative evaluation, we have computed the two metrics between pseudo-labels and GT labels, including MAE (5.124) and RMSE (6.687). We will incorporate these results into final version for quantitative analysis.

For Question 2

The pseudo-labeling process does not significantly increase training time, primarily due to two key innovations:

• While the label propagation is an iterative process (Eq. 5), we prove in Assertion 3.1 and Eq. 6 that its convergent solution can be obtained through simple matrix operations (Eqs. 6-7), eliminating the need for actual iterations.
• For constructing the nearest neighbor graph matrix (Sec 3.3.2), we employ FAISS (as stated in Right Line 303) - an efficient dense vector similarity search library developed by Facebook AI Research, which dramatically accelerates the pseudo-label learning process.

To further validate the efficiency of pseudo-label learning, we tested on a single 2080 Ti GPU using SPAQ with the same experimental setup as in Table 3 (SPAQ 1:8:1). The results show that: (1) Stage 1 training required 63.24 seconds per epoch (5 epochs in total); (2) In Stage 2, pseudo-label generation took 19.38 seconds while model training required 918.56 seconds per epoch (5 epochs in total). This demonstrates that the pseudo-labeling process accounts for only 2.1% of Stage 2 training time and 1.9% of the combined runtime for both stages.

For Question 3

Indeed, incorporating additional unlabeled data enhances model performance, and the same holds true for labeled data. While the results in Tables 5 & 10 partially support this observation, we conducted the following supplementary experiments to further validate the conclusion:

On the one hand, to systematically validate this effect on KonIQ-10k, we varied the proportion of unlabeled data ($N_u$) following the settings in Table 1:

On the other hand, following Table 11's settings, we only increased the labeled training data proportion ($N_l$) from KonIQ-10k while using SPAQ as unlabeled training data:

The above results and Tables 5 & 10 can confirm that expanding labeled or unlabeled datasets can both improve model performance.

For Question 4

The key limitation of our algorithm stems from the label conversion module based on entropy minimization, which inherently presumes high confidence in the annotated quality score distributions. In practice, annotation biases (such as significant labeling noise or uneven MOS distributions) may compromise pseudo-label reliability. For example, when adding Gaussian noise (μ=0.5, σ=0.5) to the MOS labels of training samples while keeping the Table 1 configuration, PLCC and SRCC decreased to 0.632 and 0.607 respectively.

Therefore, our future work will focus on developing noise-robust BIQA techniques. We will further add and analyze these limitations in final version.