DDR: Exploiting Deep Degradation Response as Flexible Image Descriptor

Image deep features extracted by pre-trained networks are known to contain rich and informative representations. In this paper, we present Deep Degradation Response (DDR), a method to quantify changes in image deep features under varying degradation conditions. Specifically, our approach facilitates flexible and adaptive degradation, enabling the controlled synthesis of image degradation through text-driven prompts. Extensive evaluations demonstrate the versatility of DDR as an image descriptor, with strong correlations observed with key image attributes such as complexity, colorfulness, sharpness, and overall quality. Moreover, we demonstrate the efficacy of DDR across a spectrum of applications. It excels as a blind image quality assessment metric, outperforming existing methodologies across multiple datasets. Additionally, DDR serves as an effective unsupervised learning objective in image restoration tasks, yielding notable advancements in image deblurring and single-image super-resolution. Our code is available at: https://github.com/eezkni/DDR

Paper

References (69)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer tuJD7/10 · confidence 4/52024-07-11

Summary

This paper proposed a low-level visual descriptor with text embeddings and explored its application on many low-level vision tasks.

Strengths

1. The experiment is sufficient and detailed. 2. The paper is well-written.

Weaknesses

**Weaknesses of the methods.** 1. This paper uses a text-based model to describe the low-level visual information of the images, which seems to be similar to Q-Bench [1]. I suggest that the author compare it with it and emphasize the differences between the method proposed and Q-Bench. 2. Simple super-resolution networks can identify different low-level visual information [2,3], such as degradation types. The author should compare with this method. Also, the description of lines 69 and 70 could be appropriately modified based on this. 3. The network settings for single-image super-resolution (SISR) are somewhat unreasonable. NAFNet and Restormer have a lot of downsampling and are not suitable as base models for SISR experiments. I suggest to use RCAN, SwinIR, and HAT. [1] Wu H, Zhang Z, Zhang E, et al. Q-Bench: A Benchmark for General-Purpose Foundation Models on Low-level Vision. [2] Liu Y, Liu A, Gu J, et al. Discovering distinctive" semantics" in super-resolution networks. [3] Liu Y, He J, Gu J, et al. Degae: A new pretraining paradigm for low-level vision.

Questions

1. The order of the figures and tables in this paper seems a little confusing and should be more carefully formatted. For example, the experimental results of SISR are not introduced until the end of page 8, but the result images are shown at the beginning of page 8. 2. The experimental settings are not fully given, such as the learning rate and training resolution, which may lead to unfair comparison. For example, SISR models are often trained on 48 $\times$ 48 resolution, are the NAFNet and Restormer also trained on 48 resolution for SISR tasks in this paper?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Please see the weaknesses and questions.

Reviewer mUCn5/10 · confidence 5/52024-07-12

Summary

The paper introduces Deep Degradation Response (DDR), a method to quantify changes in image deep features under varying degradation conditions. DDR facilitates flexible and adaptive degradation through text-driven prompts. It reports to excel in blind image quality assessment and image restoration tasks like deblurring and super-resolution. The paper compares the proposed DDR with the existing techniques across multiple datasets. The authors plan to release their code for public use.

Strengths

1. DDR demonstrates effectiveness across multiple applications, including Blind Image Quality Assessment (BIQA) and image restoration. Its adaptability to different degradation scenarios is a significant advantage. 2. The text-driven approach allows DDR to adjust degradation levels based on specific requirements, making it versatile and applicable to various use cases.

Weaknesses

1. The performance on BIQA is not particularly competitive, and some of the latest IQA metrics, such as LIQE, UNIQUE, and TreS, are not included for comparison. Please refer to the paper “Blind Image Quality Assessment via Vision-Language Correspondence: A Multitask Learning Perspective”. 2. In Equation 8, DDR is calculated using the restored image and its corresponding degradation. According to the results presented in Table 4, the proposed DDR performs better than PSNR+LPIPS. Since LPIPS directly minimizes the feature difference between the restored image and the original image, it should theoretically be more effective for image restoration. The paper does not explain why DDR outperforms the LPIPS loss. Moreover, in Table 6, the pre-trained model also influences the performance of DDR. Does this imply that the gains achieved by DDR might be due to a stronger backbone in comparison to LPIPS? 3. The paper utilizes DDR as an image quality assessment metric but does not provide an explanation for why it could represent the quality of an image.

Questions

Given that the negative and positive values are fixed, T_d)is a constant vector for each degradation. DDR seems to measure the disparity after adding \hat{T_d} to the feature. Does a high DDR indicate that the image distribution is robust to the feature interference caused by T_d?

Rating

5

Confidence

5

Soundness

2

Presentation

2

Contribution

2

Limitations

The authors address the limitations and potential societal impact of their work.

Reviewer mUCn2024-08-12

Thanks for authors' response on the comments. Most of my concerns have been addressed.

Authorsrebuttal2024-08-12

Thanks for your comments

Thank you very much for your thoughtful comments and appreciation of our work. We will do our best to improve the final version of our paper based on your valuable suggestions.

Reviewer Y3EC5/10 · confidence 4/52024-07-13

Summary

In this paper, the authors propose a feature descriptor to assess low-level image quality degradations. Based on CLIP, the proposed method first encodes input image and its degraded version to features in CLIP space; the input image is encoded by CLIP image encoder, and the degraded image feature is generated by adding a textual feature of degradation to the image feature. The Deep Degradation Response (DDR) is measured by calculating the distance (seems that cosine is used) of two features. The authors demonstrated the effectiveness of the proposed descriptor with extensive experiments and analysis including SRCC test and applications to image deblurring and super-resolution.

Strengths

1. The paper proposes to exploit CLIP feature to measure image quality under various degradations. Unlike other image quality works focusing on image-based approaches, this paper introduces a novel approach using textual features. 2. Surprisingly, it seems that the proposed method works well without training CLIP with degradation prompts (I need clarification of it in Weakness). It demonstrates that CLIP can be used as a tool for image quality assessment.

Weaknesses

1. Clarity: Some technical details are unclear, which limited the understanding of the proposed method. - I did not understand the distribution of DDR in Figure 2. Why is the distribution of "adaptive" better than "low" and "high"? What did the authors do specifically for "optimal"? - The authors mention in L84-85 that there are options in the disparity metric M, but in SRCC evaluation there exist a clear positive or negative direction of correlation. Since Ln and cosine metrics have different meaning for low/high values, more clarification is needed in the description of the method. It seems that the authors used cosine metric in their code. - It seems that the method simply used pre-trained CLIP features for both images and texts. Is it correct? - Among degradation types, Color and Content seem ambiguous. Do the authors have a clear definition of these categories? 2. Even though the DDR can score good and bad quality images in terms of (color, noise, blur, ...), it seems difficult to measure the amount of degradation such as noise level. It makes me wonder if the DDR contains useful information to measure degradations, or simply focus on visually pleasing images because positive texts usually correspond to those images? This is a reasonable question as the authors are aware of the fact that the CLIP features may have bias to high-level features rather than low-level degradations. 3. Continuing 2., although the authors framed the paper as image descriptor for degradations, I think the paper is more relevant to non-reference image quality assessment. Therefore, more prior works in this line of research need to be discussed in Sec. 2. 4. Although it is mentioned in the limitation, the proposed method is evaluated on one type of sentence per each degradation. Can the authors justify why they chose these specific words? Due to the nature of texts, there would be similar words that share similar meanings. How robust is the proposed method against the choice of words?

Questions

The paper generally addresses an interesting problem of image quality descriptor using CLIP textual features, and the results look encouraging. However, it lacks the in-depth understanding of the CLIP and proposed method. I would like to hear the authors' answers to the points raised in Weaknesses before adjusting my rating.

Rating

5

Confidence

4

Soundness

2

Presentation

2

Contribution

3

Limitations

The authors described two limitations of the proposed method. A potential negative societal impact would be - The proposed method could be potentially used to discriminate a certain group of photos (e.g. race, gender, etc) by including those words in degradation prompts.

Reviewer tuJD2024-08-12

Thanks for authors' response on the comments. I still have some concerns about DDR. **Comparison with Q-bench.** Could we think that Q-Bench explicitly aligns the two modalities of degradation and text, while DDR aligns these two modalities in latent space?

Authorsrebuttal2024-08-12

We appreciate for your thoughtful comments and insights. Regarding your concern towards the comparison with Q-bench, we acknowledge that both Q-bench and the proposed DDR approach necessitate alignment between modalities of degradation and text. Specifically, Q-bench explicitly instructs MLLMs to align these modalities, while DDR requires alignment in latent space to facilitate the generation of degradation representations. Your insights have provided a valuable perspective on this issue.

Reviewer tuJD2024-08-12

This difference shows that this paper contributed to the IQA field. My concerns have been addressed and I will raise the score to accept (7).

Authorsrebuttal2024-08-12

Thanks for your comments

Thank you very much for your thoughtful comments and appreciation of our work. We will do our best to improve the final version of our paper based on your valuable suggestions.

Reviewer Y3EC2024-08-13

Thank the authors for addressing the raised concerns. I have read all reviews and the authors' responses. All of my concerns are resolved and I think the paper has a valid contribution to be published at the conference. I will increase my rating to borderline accept.

Authorsrebuttal2024-08-13

Thanks for your comments

Thank you very much for your thoughtful comments and appreciation of our work. We will do our best to improve the final version of our paper based on your valuable suggestions. Thank you!

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC