NeurIPS 2020

CLEARER: Multi-Scale Neural Architecture Search for Image Restoration


Review 1

Summary and Contributions: This paper presents a method to search a multi-scale network for image restoration using the Neural Architecture Search (NAS). Moreover, the authors propose a loss function to relax the discrete architecture parameters into a continuous distribution. Experiments are performed for image denoising and image deraining.

Strengths: The paper is easy to follow. The proposed method is faster than competing NAS based image restoration approaches.

Weaknesses: 1: Limited novelty: CLEARER uses multi-scale search space that consists of three types of modules: parallel module, transition module, and fusion module. All of these modules were originally proposed in [2, 1].The authors did not cite these works when mentioning the said modules throughout the paper. 2: The proposed method generates two different architectures: one for image denoising and the other for image deraining. It seems inconvenient, as for every new task we would have a different architecture. it would have been more desirable to aim for a universal network that can be trained for different image restoration tasks. 3: The authors claim several times that CLEARER is particularly designed for image restoration. However, they did not provide any analysis/insights of what makes it specific for image restoration. For instance, what makes it suitable for image denoising and image deraining, OR why it would not work for any other applications such as semantic segmentation? 4: Experiments aren’t comprehensive. Only one dataset is used for image deraining evaluation. The reviewer would recommend testing on other commonly used image deraining datasets, such as Rain100L, Rain100H, Rain1200,etc. And for image denoising, the reviewer would recommend testing on real image datasets, such as SIDD and DND. 5: It seems that the architecture generated by CLEARER is driven by the dataset on which it is trained on, rather than the particular image restoration task. The reviewer would suggest to search for an architecture for image deraining using CLEARER, but on a new dataset (not Rain800), and see if it yields the same architecture as in Fig. 4 of the supplementary material. [1] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5693–5703, 2019. [2] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, YangZhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 2020.

Correctness: Probably correct.

Clarity: The paper is easy to follow.

Relation to Prior Work: Several references are missing. Please see "Weaknesses".

Reproducibility: Yes

Additional Feedback: Please see "Weaknesses". The issues rasied in "Weaknesses" are desired to be addressed in the rebuttal. Post Rebuttal Comments: The reviewer would first like to thank the authors for addressing the concerns raised in the initial review. The rebuttal has partially addressed the concerns raised in the initial review. However, the proposed method still has some limitations. The PSNR (30.42) and SSIM (14 0.9056) of CLEARER reported on RAIN1400 is inferior to SOTA. The same is also true for the deraining results reported for Rain800 dataset (line 4-5 in the feedback document). Further, even when changing from one deraining dataset to another, the architecture is no longer universaral and changes (line 4-5 and line 12-14). Despite the aforementioned limitations, considering NAS for image restoration tasks has received little attention. This paper is likely to encourage further work in this direction. Therefore, the reviewer would like to raise the final score from 5 to 6.


Review 2

Summary and Contributions: This paper proposes a multi-resolution neural architecture search space with three task-flexible modules for image restoration. Besides, a new loss function is used to balance the model complexity and performance, which is highly expected to a wide range of applications.

Strengths: 1. This paper employs a novel loss function to balance the model complexity and performance that is highly expected to the resource-constrained scenarios. 2. This paper employs a differentiable strategy to search when to fuse multi-resolution features. 3. Comparisons on image denoising and deraining with state-of-the-art approaches show the effectiveness of the proposed CLEARER.

Weaknesses: 1. I am curious that if the proposed method can deal with other ill-posed image restoration tasks, like image deblurring. The authors may want to discuss some of this. 2. How about exchange the block sequence of the parallel module and the transition module? 3. Iterative optimization of the mixing architecture parameters and the network weights seems to be a crucial step. How it influences the final results? 4. The recent work show that some interactions between scales promote the performance of image restoration. Therefore, can the proposed method involves some interactions between scales?

Correctness: Overall, the claims and method are easy to follow, the authors have clearly explained the proposed method.

Clarity: yes, this paper is well written and organized.

Relation to Prior Work: Yes, the authors have clearly discussed the difference, improvements between the proposed method and previous contributions.

Reproducibility: Yes

Additional Feedback: see the weakness == After rebuttal == I am satisfied with the rebuttal, and have no objection for accepting this paper.


Review 3

Summary and Contributions: This paper proposed to solve the image restoration problem in a multi-scale manner, while the model is obtained via network architecture search (NAS), with a structure space constraint loss and a computation amount loss. Extra (lower) scales are incorporated one by one in the pipeline via a transition module. Both quantitative and qualitative results are impressive.

Strengths: + The paper introduces a structure space constraint loss and a computation amount loss, and reports impressive results

Weaknesses: - The contribution in network architecture search is insufficient. - After the first “Fusion module” in the “Cell module”, the features from different scales have been mixed up. Do the authors refer to “multi-scale” as the Parallel module? Or please make it clear if there are any insights. - The loss $\scriptL_{Arch}$ in Eqn. (4) is contrary to the description in the paper, see “Correctness” for more details. - Besides, there are some minor problems. a) Is the structure of the “Transition module” fixed or searched? b) The English should be further polished, and there are some typos. E.g., overcame -> overcome in line 44, consists -> consist in line 141, constraint->constrain in line 160, etc.

Correctness: Although the authors report impressive results, there are technical errors in this paper. As stated in lines 171 and 172, the authors say that “This regularization will enforce $\alpha$ to distribute to approach either 0 or 1”. However, the loss proposed in Eqn. (4) will constrain the value of $\alpha$ to be 0.5, which is contrary to the statement.

Clarity: The English should be further polished. See the weakness.

Relation to Prior Work: Yes

Reproducibility: No

Additional Feedback: The clarification of the loss term and multi-scale have partially addressed my concerns. Considering that NAS for image restoration remains less investigated, I slightly raised my score to encourage this line of work.


Review 4

Summary and Contributions: This paper proposed a multi-scale NAS method for image denoising and deraining. This paper designed three modules and investigated the relationships of these modules. From my viewpoint, this paper substituted for the cells in general NAS methods by the modules for image restorations. In brief, this paper investigated the combination of different designed modules rather than detailed construction of inner cells. From the aspect of design philosophy, this paper introduced three basic modules, i.e., Parallel module, Transition module and Fusion module, which utilized the multi-scale feature representation and multi-resolution features integration and fusion. From the aspect of search strategies, this paper depicted the principles of loss design to reduce the model complexity.

Strengths: The proposed method is relatively similar to DARTS, which is a differential NAS method for high-level vision. Essentially speaking, it is also a cell-searching-based and gradient-based NAS method. In my opinion, it is not a very significant and insightful idea for image restoration. The main contributions are the trade-off training strategy and the loss design strategy.

Weaknesses: 1. For the principles for designed modules, this paper proposed three basic modules for interior image restoration, however, the interior structure of these modules is fixed, which is not so convinced to build these modules. In my opinion, an inner NAS strategy is necessary to search for a considerable structure for image restoration. Furthermore, as shown in Fig.1(c), the global flow is fixed, which consists of two parallel modules and the multiply combination of transition modules and cells. I wonder that the reason that not searching the global connections and demonstration of reasonability. Indeed, theses principled modules are common principles for other vision problems, such as image segmentation and optical flow estimation. Maybe other principled modules, such as attention modules and tasks-specific modules are vital to introduce. 2. From the experiment of this paper, some experiment settings are not very rational. As a significant contribution, the trade-off between model complexity and inference accuracy is not obviously shown when comparing with other methods for image denoising and deraining. In other words, I cannot find the final schemes for these tasks under the consideration of the trade-off between model complexity and inference accuracy. From the supplemented materials, the designed network is very deep. This paper should indicate the parameters of this paper and other networks. From the reproduction aspect, I cannot find how to search for a good structure for deraining. The various rain scenarios are also necessary to demonstrate your performance. Moreover, this paper aims to remove signal-dependent or -independent noises. Some experiments are missing, such as the performance on the real scenarios and RGB colorful images rather than deraining scenarios. 3. As a bilevel optimization method, it is not very clear the relationship between three losses and objective functions (upper and lower formulations). L_{arch} and L_{Comp} should be optimized in the upper functions. 4. Some subjective statements are inappropriate to introduce this paper. Some proofs and references are needed to demonstrate your statement. it is labor-intensive to seek an effective architecture, while the image recovery performance is sensitive to the choice of neural architecture. One more daunting task of multi-scale architecture design is unknown is that when to fuse the multi-scale feature. Besides these explicit multi-scale methods, the models with skip connections [10] could also be regarded as using multi-scale information in an implicit way. (The author should provide a detailed explanation to verify these statements.)

Correctness: As a bilevel optimization method, it is not very clear the relationship between three losses and objective functions(upper and lower formulations).

Clarity: The related works should use the simple past tense.

Relation to Prior Work: Yes

Reproducibility: Yes

Additional Feedback: ---------------------------------------------------------------- After seeing the rebuttal and comments from other reviewers, I agree that the main contribution of this work should not be the bilevel optimization methodology, but the new search space and training strategies for image restoration tasks. From this point of view, I think this paper should have protentional influence for NAS methods in low-level vision areas. Thus, I decide to provide a positive final recommendation for this paper.