Source-linked AI summary
SDWNet: A Straight Dilated Network with Wavelet Transformation for Image Deblurring
Wenbin Zou, Mingchao Jiang, Yunchen Zhang, Liang Chen, Zhiyong Lu, Yi Wu
TL;DR
Image deblurring is an ill-posed recovery problem, and repeated sampling can lose texture while complex multistage designs can slow convergence. SDWNet combines dilated convolution with wavelet reconstruction, and the authors report state-of-the-art performance and clearer recovered images.
Problem
Image deblurring must recover a clear image from a blurred input despite the blurring process being one-to-many, while repeated sampling in common architectures can lose texture details.
Source: §1 · pp. 1–2Method
SDWNet uses a simple CNN with different-rate dilated convolutions for varied receptive fields and wavelet reconstruction that supplies frequency information to spatial reconstruction.
Source: §1 · p. 2, §3.1 · p. 3, §3.3 · p. 4Results
Takeaways & Limitations
The authors describe SDWNet as simple in structure and easily transferable to other high-level tasks.
Source: §5 · p. 8
Abstract
from arXiv · showhide
Image deblurring is a classical computer vision problem that aims to recover a sharp image from a blurred image. To solve this problem, existing methods apply the Encode-Decode architecture to design the complex networks to make a good performance. However, most of these methods use repeated up-sampling and down-sampling structures to expand the receptive field, which results in texture information loss during the sampling process and some of them design the multiple stages that lead to difficulties with convergence. Therefore, our model uses dilated convolution to enable the obtainment of the large receptive field with high spatial resolution. Through making full use of the different receptive fields, our method can achieve better performance. On this basis, we reduce the number of up-sampling and down-sampling and design a simple network structure. Besides, we propose a novel module using the wavelet transform, which effectively helps the network to recover clear high-frequency texture details. Qualitative and quantitative evaluations of real and synthetic datasets show that our deblurring method is comparable to existing algorithms in terms of performance with much lower training requirements. The source code and pre-trained models are available at https://github.com/FlyEgle/SDWNet.
1. Introduction
Image deblurring is an ill-posed problem, and the authors identify texture loss from repeated sampling and slow convergence in complex coarse-to-fine designs as challenges. They propose SDWNet, combining dilated convolution and wavelet reconstruction, and report state-of-the-art performance.
- Repeated upsampling and down-sampling in Encode-Decode networks can lose texture details, while coarse-to-fine designs can be complex, computationally intensive, and slow to converge.Source: §1 · p. 2
- The authors report that SDWNet achieves state-of-the-art performance with assistance from its dilated convolution and wavelet reconstruction modules.Source: §1 · p. 2
- Different dilated rates provide different receptive fields, helping the network capture non-local similar features without relying on repeated upsampling and down-sampling.Source: §1 · p. 2
- The wavelet reconstruction module uses frequency-domain information to complement the spatial domain and recover more high-frequency details.Source: §1 · p. 2
- SDWNet uses a simple, streamlined CNN structure rather than a coarse-to-fine design, which the authors say addresses difficult training and slow convergence.Source: §1 · p. 2
2. Related Work
Related work covers CNN deblurring, dilated convolution, and wavelet-based restoration. The paper positions SDWNet's different-rate dilated blocks and wavelet reconstruction against limitations in prior approaches.
2.1. Deep Image Deblurring
- Most reviewed deblurring networks use alternating down-sampling and upsampling, which can lose image information; the authors use different dilation rates to obtain varied receptive fields.Source: §2.1 · p. 3
2.2. Dilated Convolution
- Prior deblurring work applied atrous blocks with different rates, but their rate pattern did not cover all pixel points; the authors adjust rates for nearly complete receptive-field coverage.Source: §2.2 · p. 3
2.3. Related Application based on WT
- Prior wavelet restoration methods mix frequency information directly, which the authors say can cause frequency interactions and wrong textures; their reconstruction module is proposed to recover a clear image.Source: §2.3 · p. 3
3. Proposed Method
SDWNet uses a plain network combining shallow feature extraction, cascaded dilated-convolution blocks, and a parallel wavelet reconstruction branch. Its modules combine different receptive fields and recover frequency subbands to preserve fine image texture.
3.1. Framework
- A 7 × 7 shallow-feature layer feeds cascaded dilated-convolution blocks and a parallel wavelet reconstruction branch in the plain SDWNet architecture.The framework predicts a residual that is added to the blurred input; training uses Charbonnier and SSIM losses.Source: §3.1 · pp. 3–4
- SDWNet predicts a residual image and adds it to the blurred input; its loss combines Charbonnier and SSIM terms.The SSIM-loss weight λ is a hyperparameter.Source: §3.1 · p. 4
3.2. Dilated Convolution Module
- DCBs use parallel convolutions at different dilation rates to fuse multi-receptive-field information, with distinct rates across the last and other layers.The last DCB layer uses rates {1, 3, 5, 7}, while other DCB layers use {1, 2, 4, 8}; the final layer's non-overlapping rates are intended to avoid gridding effects.Source: §3.2 · p. 4, Figure 3 · p. 4
3.3. Wavelet Reconstruction Module
- The wavelet reconstruction module decomposes input features into four frequency subbands, recovers them separately with three-layer convolution, then applies the inverse transform.Separate subband recovery is used to avoid interference between different frequency subbands.Source: §3.3 · pp. 4–5, Figure 4 · p. 5
4. Experiments with Analysis
Experiments evaluate SDWNet on synthetic and real deblurring datasets, with sliding-window crops and a training setup requiring no pretraining. Results report strong benchmark performance, efficiency, qualitative detail recovery, and ablation findings.
4.1. DataSet
- The GoPro dataset contains 3,214 pairs, split into 2,103 training and 1,111 evaluation pairs; HIDE is synthetically generated, while RealBlur uses real-world captures.RealBlur-J uses camera JPEG outputs, and RealBlur-R is processed from RAW images.Source: §4.1 · p. 5
4.2. Dataset Sliding Crop
- Sliding-window cropping uses 480 × 480 patches with a 240-pixel step and edge compensation, yielding 24 patches per image and up to 50,472 patches.GoPro images are 1280 × 720; Figure 5 depicts the crop blocks and edge compensation.Source: §4.2 · p. 5, Figure 5 · p. 5
4.3. Implementation Details
- SDWNet is end-to-end trainable without pretraining; training uses 1,500 GoPro epochs followed by 50 epochs on cropped GoPro data.The implementation uses AdamW, input size 416 × 416, batch size 8, and λ = 1.Source: §4.3 · p. 5
4.4. Image Deblurring Results
- Compared with DMPHN, SDWNet improves PSNR by 0.16dB and SSIM by 0.027, while achieving the best results on GoPro and HIDE.Table 1 compares PSNR (dB) and SSIM on GoPro and HIDE.Source: §4.4 · pp. 6–7, Table 1 · p. 7
- On RealBlur-R, SDWNet leads previous methods by 2.51dB PSNR; on RealBlur-J, its PSNR and SSIM are similar to previous best methods.Table 2 covers direct application of the GoPro-trained model and RealBlur training/testing.Source: §4.4 · p. 7, Table 2 · p. 7
- Visual comparisons report that SDWNet recovers clear images where older methods remain blurry and recent methods produce erroneous textures.The authors attribute correct detail recovery to dilated convolution and high-frequency detail recovery to wavelet transformation.Source: Figure 6 · p. 6, §4.4 · p. 7
- SDWNet is the fastest method in Table 3 and achieves better PSNR and SSIM than DMPHN with one-third of its parameters and FLOPs.The table reports GoPro test-set runtimes on an Nvidia Titan Xp GPU and FLOPs for 256 × 256 inputs.Source: §4.4 · p. 7, Table 3 · p. 8
4.5. Ablation Studies
- Ablations select d = 16 and w = 32, favor ELU and bilinear upsampling, report gains from the adjusted dilation rate and WRM, and find best performance at λ = 1.The authors report width has a greater effect than depth; Figure 7 studies λ values from 0 to 2.Source: Figure 7 · p. 7, §4.5 · p. 8, Table 4 · p. 8
5. Conclusion
The proposed dilated-convolution network combines different receptive fields with wavelet-based high-frequency information for image deblurring. The authors report clearer restorations than other methods and describe the structure as simple and transferable to other high-level tasks.
5. Conclusion
- Qualitative and quantitative results show that the algorithm restores clearer images than other methods.The authors also describe the approach as simple in structure.Source: §5 · p. 8
- Dilated convolutions at different rates provide information from different receptive fields, while wavelet transformation supplies high-frequency information for spatial-domain reconstruction.The authors say the wavelet information results in clearer images.Source: §5 · p. 8
- The authors propose exploring the approach for other image restoration tasks, including denoising and super-resolution.They describe the network structure as easily transferable to other high-level tasks.Source: §5 · p. 8