Source-linked AI summary
Multi-stage image denoising with the wavelet transform
Chunwei Tian, Menghua Zheng, Wangmeng Zuo, Bob Zhang, Yanning Zhang, David Zhang
TL;DR
Existing CNN denoisers often increase network depth to improve denoising, which may make training difficult. MWDCNN uses dynamic convolution, cascaded wavelet-enhancement blocks, and residual refinement; it reportedly outperforms DnCNN and ADNet in quantitative and qualitative analyses. The method is also described as beneficial for damaged images with unknown noise, but it requires supervised training.
Problem
Most existing CNN denoisers rely on increasing network depth for better performance, which may cause training difficulty.
Method
MWDCNN is a three-stage denoising CNN combining dynamic convolution, cascaded wavelet-enhancement blocks, and an enhanced residual dense architecture.
Results
MWDCNN outperforms DnCNN and ADNet in quantitative and qualitative image-denoising analysis.
Takeaways & Limitations
The proposed method is described as beneficial for damaged images with unknown noise.
Takeaways & Limitations
The denoising model depends on supervised training, and noisy camera images are used in the training setting.
Abstract
from arXiv · showhide
Deep convolutional neural networks (CNNs) are used for image denoising via automatically mining accurate structure information. However, most of existing CNNs depend on enlarging depth of designed networks to obtain better denoising performance, which may cause training difficulty. In this paper, we propose a multi-stage image denoising CNN with the wavelet transform (MWDCNN) via three stages, i.e., a dynamic convolutional block (DCB), two cascaded wavelet transform and enhancement blocks (WEBs) and a residual block (RB). DCB uses a dynamic convolution to dynamically adjust parameters of several convolutions for making a tradeoff between denoising performance and computational costs. WEB uses a combination of signal processing technique (i.e., wavelet transformation) and discriminative learning to suppress noise for recovering more detailed information in image denoising. To further remove redundant features, RB is used to refine obtained features for improving denoising effects and reconstruct clean images via improved residual dense architectures. Experimental results show that the proposed MWDCNN outperforms some popular denoising methods in terms of quantitative and qualitative analysis. Codes are available at https://github.com/hellloxiaotian/MWDCNN.
1. Introduction
Image denoising methods have progressed from manually designed and optimized approaches to CNN-based discriminative learning, but deeper CNNs can make training difficult. MWDCNN addresses this with dynamic convolution, cascaded wavelet-enhancement blocks, and residual refinement, and reportedly outperforms DnCNN and ADNet quantitatively and qualitatively.
- Earlier denoising methods manually selected parameters and could require complex optimization algorithms for strong denoising effects.
- CNN-based denoising mines image structure information through discriminative learning rather than manually tuning parameters.
- Existing effective denoising methods may rely on deeper architectures, increasing training difficulty.
- MWDCNN combines a dynamic convolutional block, two cascaded wavelet transform and enhancement blocks, and a residual block.The dynamic convolution adjusts convolution parameters, while the wavelet and residual components support denoising and feature refinement.
- MWDCNN reportedly outperforms DnCNN and ADNet in quantitative and qualitative image-denoising analysis.
2. Related work
Related work combines CNNs with adaptive convolutions and signal-processing techniques to improve denoising while managing computational demands. Wavelet transforms provide frequency information that can complement CNN-derived structural information for low-level vision tasks.
- Many CNN denoisers use shared convolutional weights for noisy images, while dynamic convolutions adjust parameters to address this limitation.Dynamic convolutions can fuse multiple parallel convolutions through attention mechanisms and have been used to reduce computational costs.
- Dynamic convolutions have been introduced into CNNs for denoising to adjust convolution-kernel parameters and train a robust denoiser.
- Wavelet transforms are used in low-level vision to filter noise and learn detailed and content information.Prior work combines wavelet transforms with statistical estimation, genetic algorithms, or CNNs for image restoration tasks.
- MWDCNN fuses wavelet-derived frequency features with CNN structural information to obtain complementary information for image denoising.
3. The proposed method
MWDCNN is a 23-layer, three-stage denoising network combining dynamic convolution, cascaded wavelet enhancement, and residual refinement. Its stages adapt convolutional parameters, fuse frequency and structural information, and reconstruct clean images through residual learning.
- Network architecture: MWDCNN contains a dynamic convolutional block, two cascaded wavelet transform and enhancement blocks, and a residual block.The network is described as a 23-layer architecture.
- Dynamic convolution block: Dynamic convolution adjusts parameters of several convolutions according to different images, trading off denoising performance and computational cost.Its weight generator produces four weights for four parallel convolutional kernels.
- Loss function: MWDCNN is trained with mean square error on supervised clean/noisy image pairs and optimized using Adam.The training set contains n noisy images paired with clean images.
- Dynamic convolution block: The dynamic convolution block uses a five-layer structure beginning with a convolutional layer and including a three-layer dynamic convolution.The dynamic convolution contains a two-layer weight generator and a 5 × 5 convolutional layer.
- Wavelet transform and enhancement blocks: Each wavelet enhancement block applies DWT, feature enhancement, and IDWT to combine frequency features with structural information and suppress noise.Feature enhancement uses a residual dense block configuration to recover more detailed information.
- Residual block: The residual block refines features with an enhanced residual dense architecture and reconstructs clean images through convolution and residual learning.It uses two residual dense block–ReLU combinations, residual operations, and a convolutional refinement step.
4. Experiments
Experiments evaluate MWDCNN on synthetic and real noisy-image datasets, including ablations of its dynamic convolution, wavelet-enhancement, residual, and loss-function components. Across quantitative and qualitative comparisons, the method is reported to offer competitive denoising performance, adaptability, and computational efficiency.
- Experimental settings: Training uses BSD-derived images for synthetic denoising and 100 natural 512 × 512 images for real-noisy denoising.Synthetic training produces 221,184 noisy 48 × 48 patches, while real-noisy training uses 211,600 patches.
- Network analysis: The dynamic convolutional block improves denoising performance while requiring competitive parameter and FLOP costs compared with six stacked convolutional layers.A three-convolution plus dynamic-convolution configuration obtains higher values than six stacked convolutions in the reported ablation.
- Network analysis: Wavelet transformation and discriminative learning improve MWDCNN performance, and two stacked WEBs are reported effective for image denoising.A model with wavelet-enhancement components outperforms variants using a convolutional layer without DCB and two WEBs, or without two WEBs.
- Network analysis: The improved residual dense architecture strengthens denoising results compared with variants lacking DCB, WEBs, or the residual design.The reported ablations support the effectiveness of the residual block and its RLO and RDB components.
- Comparisons with state-of-the-art denoising methods: Qualitative comparisons report clearer observation areas for MWDCNN than other denoising methods, supporting its stronger visual denoising results.The paper summarizes MWDCNN as effective according to quantitative and qualitative analysis.
5. Conclusion
MWDCNN combines adaptive convolution, wavelet and residual-dense feature processing, and improved residual architectures in a three-stage denoising CNN. The method is useful for unknown-noise images but depends on supervised training with clean references.
- Conclusion: MWDCNN uses three stages: dynamic convolution, wavelet-transform enhancement, and improved residual-dense refinement.The stages dynamically adjust convolution parameters, fuse frequency and structure features, and remove redundant features.
- Conclusion: Dynamic convolution linearly combines several kernels to adjust parameters for different noisy images, trading denoising performance against computational cost.The paper states this design is useful for noisy images with unknown noise.
- Conclusion: Wavelet transformation and residual dense blocks fuse frequency and structure features to suppress noise and enhance robustness in complex scenes.This combination is intended to recover more useful image information while improving denoiser robustness.
- Conclusion: Improved residual dense architectures remove redundant features, improving denoising performance and supporting clean-image reconstruction.The residual block refines obtained features before reconstruction.
- Conclusion: The method benefits damaged images with unknown noise but depends on supervised training, while camera-collected noisy images make clean references difficult to obtain.The authors propose using clean reference images to train a blind denoiser in future work.