Source-linked AI summary
Dual Residual Attention Network for Image Denoising
Wencong Wu, Shijie Liu, Yi Zhou, Yungang Zhang, Yu Xiang
TL;DR
Real-world, spatially variant noise remains difficult for many CNN denoisers, limiting practical image-denoising applications. DRANet combines two complementary residual-attention branches with feature filtering and enlarged receptive fields, and experiments show competitive performance on synthetic and real-world noise.
Problem
Many CNN denoisers perform well on spatially invariant noise but struggle with real, spatially variant noise encountered during image acquisition or transmission.
Method
DRANet uses two parallel branches with RAB and HDRAB blocks, skip connections, feature fusion, downsampling, and dilated convolutions to learn complementary local and global features.
Results
DRANet achieves competitive denoising performance across different denoising tasks, including synthetic and real-world noise removal.
Takeaways & Limitations
DRANet offers an attractive balance between network complexity and denoising performance for practical image denoising tasks.
Abstract
from arXiv · showhide
In image denoising, deep convolutional neural networks (CNNs) can obtain favorable performance on removing spatially invariant noise. However, many of these networks cannot perform well on removing the real noise (i.e. spatially variant noise) generated during image acquisition or transmission, which severely sets back their application in practical image denoising tasks. Instead of continuously increasing the network depth, many researchers have revealed that expanding the width of networks can also be a useful way to improve model performance. It also has been verified that feature filtering can promote the learning ability of the models. Therefore, in this paper, we propose a novel Dual-branch Residual Attention Network (DRANet) for image denoising, which has both the merits of a wide model architecture and attention-guided feature learning. The proposed DRANet includes two different parallel branches, which can capture complementary features to enhance the learning ability of the model. We designed a new residual attention block (RAB) and a novel hybrid dilated residual attention block (HDRAB) for the upper and the lower branches, respectively. The RAB and HDRAB can capture rich local features through multiple skip connections between different convolutional layers, and the unimportant features are dropped by the residual attention modules. Meanwhile, the long skip connections in each branch, and the global feature fusion between the two parallel branches can capture the global features as well. Moreover, the proposed DRANet uses downsampling operations and dilated convolutions to increase the size of the receptive field, which can enable DRANet to capture more image context information. Extensive experiments demonstrate that compared with other state-of-the-art denoising methods, our DRANet can produce competitive denoising performance both on synthetic and real-world noise removal.
1. Introduction
Image denoising methods face a trade-off between practical applicability and representation capacity: traditional optimization methods can be slow and unstable, while deeper CNNs increase parameters and inference time. DRANet addresses these issues with a dual residual-attention architecture designed for synthetic and real-world noise, achieving competitive denoising results.
- Traditional denoising models can achieve good results but suffer from long inference times and unstable performance caused by complex optimization and many hyperparameters.
- DNN-based denoising models generally use fewer hyperparameters and shorter inference times than traditional methods.
- Increasing network depth improves learning ability but also raises parameter counts and inference time, motivating wider architectures for practical denoising.
- Attention-guided models use feature filtering to improve denoising, while skip and dense connections further enhance feature learning.
- DRANet uses two parallel branches with RABs and HDRABs to learn complementary features, filter unimportant features, and capture local and global information.
- DRANet is designed for blind denoising on synthetic and real-world noise, with a simpler architecture and fewer parameters than many state-of-the-art blind denoising models.
- Extensive experiments report competitive denoising results compared with other state-of-the-art methods.
2. Related work
Related work develops blind denoising, residual, dilated, attention-based, and dual-network approaches to improve flexibility, feature extraction, receptive fields, and denoising performance. These methods motivate combining dual branches with attention in DRANet.
- Many blind denoising models estimate noise levels from noisy images to avoid manually preset noise-level inputs, although accurate estimation remains challenging.
- DnCNN and RIDNet apply residual learning, batch normalization, or attention modules to improve denoising performance and accelerate training.
- Residual learning helps address degradation in deeper networks and mitigates vanishing or exploding gradients in image restoration.
- Dilated convolutions enlarge receptive fields without increasing network complexity, while hybrid dilated filters address their gridding phenomenon.
- Attention mechanisms are used to extract informative features for image processing and denoising, including noise-aware and feature-attention designs.
- Dual denoising networks use parallel branches to extract structures and details or combine local and global features, with later work exploring dual CNNs and attention together.
3. The proposed model
DRANet uses two complementary residual-attention branches to learn local and global image features for denoising. Downsampling, dilated convolutions, skip connections, feature fusion, and tailored losses support multiscale feature extraction and reconstruction.
- Network architecture: DRANet contains upper and lower sub-networks with five RABs and five HDRABs, respectively, whose features are fused by concatenation.Long skip connections connect blocks within each branch, while the two branches provide complementary features.
- Network architecture: The upper branch uses strided and transposed convolutions for downsampling and upsampling, extracting multiscale features and increasing the receptive field.The architecture uses 2 × 2 strided convolutions and 2 × 2 transposed convolutions.
- Residual attention block: RAB combines residual convolutions with spatial attention to extract local features, learn global context, emphasize useful spatial features, and filter uninformative features.Its skip connections fuse features between convolutional layers, while pooling and attention operations represent whole-image statistics.
- Hybrid dilated residual attention block: HDRAB combines hybrid dilated residual convolutions with channel attention to capture local features and model inter-channel relationships.Its dilated rates range from 1 to 4, and the channel attention module uses GAP, convolution, ReLU, and Sigmoid operations.
- Loss function: Training uses MSE for Gaussian denoising, whereas real-noise training combines Charbonnier reconstruction loss with edge loss to preserve high-frequency information.The edge-loss weight is empirically set to 0.1, and the Charbonnier constant ϵ is set to 10^-3.
4. Experiments and results
Experiments evaluate DRANet on synthetic and real noise using multiple datasets, ablations, visual comparisons, and efficiency analyses. DRANet generally achieves strong denoising performance, including leading results across several synthetic settings and competitive real-noise results.
- Experimental setup: Seven public datasets evaluate DRANet on synthetic and real noise, including Flick2K, SIDD, BSD68, Set12, CBSD68, Kodak24, and McMaster.Synthetic experiments use Flick2K for training and multiple grayscale and color datasets for evaluation, while real-noise experiments use SIDD and DND.
- Ablation study: The complete DRANet achieves the best ablation performance, exceeding the upper and lower branches by 0.75 dB and 0.13 dB in average PSNR.Long skip connections and residual learning slightly improve denoising performance but accelerate training and convergence.
- Grayscale synthetic noise: On Set12, DRANet has competitive average PSNR at noise level 15 and the top average PSNR results at noise levels 25 and 50.The evaluation compares multiple denoising methods across three AWGN noise levels.
- Grayscale synthetic noise: On Set12, DRANet matches BRDNet and RIDNet in average SSIM at noise level 15 and outperforms all other methods at levels 25 and 50.The reported SSIM comparison covers the same three noise levels.
- Color synthetic noise: DRANet leads average PSNR and SSIM at all tested noise levels across the CBSD68, Kodak24, and McMaster color-image datasets.The visual comparison also reports detail loss or over-smoothing for several competing methods, while DRANet reconstructs high-quality images.
- Real noise removal: On real noise, DRANet achieves very competitive results on SIDD and DND, with leading PSNR and SSIM in the reported visual comparison.The real-noise evaluation compares DRANet with classical, CNN-based, and other blind denoising models.
- Efficiency and complexity: DRANet offers competitive and faster inference than several methods, but runs longer than others while achieving better denoising performance than those faster competitors.The cited comparison includes both traditional and neural denoising methods.
5. Conclusion and future work
DRANet combines parallel residual-attention sub-networks, skip connections, feature fusion, and residual learning to improve feature extraction and denoising. The authors report competitive performance across denoising tasks while balancing network complexity and performance for practical use.
- Conclusion: DRANet uses two parallel sub-networks with different structures to learn complementary image features and select appropriate features.Residual attention blocks, long skip connections, and global feature fusion support local and global feature learning.
- Conclusion: Long skip connections and global feature fusion help DRANet obtain global image features, while residual learning accelerates training.The architecture combines branch-level skip connections with fusion between the two sub-networks.
- Conclusion: DRANet achieves competitive denoising performance on different denoising tasks while balancing network complexity and performance.The authors identify this balance as making DRANet an option for practical image denoising tasks.
- Conclusion: The paper reports parameter counts for different denoising models to support comparison of network complexity.Table 9 lists model parameter numbers in thousands.