Source-linked AI summary
Real Image Denoising with Feature Attention
Saeed Anwar, Nick Barnes
TL;DR
Existing CNN denoisers perform better on spatially invariant synthetic noise than on real noisy photographs, while practical real-image methods may require multiple stages. RIDNet uses a single-stage modular CNN with residual-on-residual connections and feature attention, and experiments report effective performance across synthetic and real-noise datasets. Its real-image evaluation is bounded by datasets whose clean references are unavailable or whose scores require an online system.
Problem
Existing denoising models have limited performance on real noisy photographs, while practical algorithms should handle varied noise with one efficient model.
Method
RIDNet is a single-stage modular CNN using residual-on-residual connections, feature attention, and a wide receptive field from dilated kernels.
Results
RIDNet achieves state-of-the-art results across three synthetic and four real-noise datasets, including a 1.17dB PSNR lead over CBDNet on a real-noise comparison.
Takeaways & Limitations
A single blind denoising network can address synthetic and real noisy photographs using the same model across the evaluated datasets.
Takeaways & Limitations
RNI15 lacks clean images, permitting only qualitative comparison, while DnD reference images are not public and PSNR/SSIM require an online system.
Abstract
from arXiv · showhide
Deep convolutional neural networks perform better on images containing spatially invariant noise (synthetic noise); however, their performance is limited on real-noisy photographs and requires multiple stage network modeling. To advance the practicability of denoising algorithms, this paper proposes a novel single-stage blind real image denoising network (RIDNet) by employing a modular architecture. We use a residual on the residual structure to ease the flow of low-frequency information and apply feature attention to exploit the channel dependencies. Furthermore, the evaluation in terms of quantitative metrics and visual quality on three synthetic and four real noisy datasets against 19 state-of-the-art algorithms demonstrate the superiority of our RIDNet.
1. Introduction
Image denoising is broadly useful, but existing methods struggle to provide one efficient model for both synthetic and real, spatially varying noise. RIDNet addresses this gap with a single-stage modular CNN using feature attention and residual connections.
- Image denoising supports image analysis, Bayesian evaluation of image priors, and unrolled solutions for broader restoration tasks.
- Model-based methods are computationally expensive and struggle with spatially variant noise and complex image textures.
- Current learning-based models improve denoising but remain limited and tailored to specific noise levels.
- RIDNet targets an efficient single model that handles synthetic and real noise without requiring separate stages.
- The model introduces feature attention as a denoising component and uses a modular design whose performance can improve by increasing the number of modules.
- Experiments cover three synthetic and four real-image noise datasets with quantitative and qualitative comparisons against state-of-the-art methods.
2. Related Works
Prior denoising research spans patch-based, probabilistic, optimization-based, and CNN approaches, while real-image methods often separate noise estimation from denoising. RIDNet instead uses an end-to-end architecture without separate subnetworks or manual intervention.
- Patch-based methods exploit self-similarity through non-local filtering, transform domains, dictionaries, and sparsity constraints.
- CSF combines half-quadratic optimization with random-field modeling and assumes a quadratic data-fidelity term with a Fourier-domain solution.
- CNN denoisers such as DnCNN and IrCNN predict image residue and improve results using repeated convolutional, normalization, and activation blocks.
- Trainable reaction-diffusion and related models can require substantial data, fine-tuning, hyperparameter selection, or stage-wise training.
- Real-image denoising is difficult because noise can be spatially variant, signal-dependent, and produced by multiple imaging and noise sources.
- CBDNet uses separate noise-estimation and non-blind-denoising subnetworks, multiple losses, and both real and synthetic-real noise training.
- RIDNet learns noise and denoises real noisy images end to end without separate subnetworks or manual intervention.
3. CNN Denoiser
RIDNet uses a three-module architecture combining residual-on-residual feature learning with feature attention. Dilated convolutions broaden the receptive field, while channel-wise gating emphasizes informative features.
- Network Architecture: RIDNet extracts initial features, processes them through residual-on-residual feature learning, and reconstructs the denoised image.Feature extraction and reconstruction each use one convolutional layer, while the central module cascades enhancement attention modules.
- Residual-on-Residual Learning: Kernel dilation in each enhancement attention module provides a wide receptive field despite the network’s small depth.The initial two branch convolutions in each module use dilation to expand contextual coverage.
- Optimization: The model optimizes a single ℓ1 loss over noisy inputs and corresponding ground-truth images.The objective minimizes the difference between RIDNet outputs and ground truth across training pairs.
- Residual-on-Residual Learning: Enhancement attention modules combine residual connections and feature attention, with four modules used in the network.Each module contains residual structure, local and short skip connections, and feature attention.
- Architectural Motivation: The study identifies feature attention as a denoising contribution and reports that residual connections ease information flow across groups.The architecture uses long, short, and local skip connections within its modular design.
- Feature Attention: Feature attention addresses unequal channel importance by modeling dependencies among channel features.Global average pooling summarizes image-wide statistics before soft-shrinkage and sigmoid-based self-gating rescales channel features.
4.1. Training settings
Training uses synthetic images from three datasets and real noisy patches from three datasets, with augmentation and Adam optimization. Evaluation uses PSNR.
- Data: Synthetic training data come from BSD500, DIV2K, and MIT-Adobe FiveK, producing 4k images.Real noisy data use 512×512 cropped patches from SSID, Poly, and RENOIR.
- Data Augmentation: Training augmentation applies random rotations of 90°, 180°, and 270° plus horizontal flipping.These transformations are applied to the training images.
- Optimization: Each training batch contains 32 input patches of size 80 × 80, optimized with Adam using an initial learning rate of 10^-4.The learning rate is halved after 10^5 iterations.
- Implementation: The network is implemented in PyTorch, trained on an Nvidia Tesla V100 GPU, and evaluated with PSNR.The passage specifies the implementation framework, hardware, and evaluation metric.
4.2. Ablation Studies
Ablation studies show that skip connections and feature attention improve denoising performance on BSD68. Removing any skip-connection type lowers performance, while greater depth without skips does not help.
- Skip Connections: The highest BSD68 PSNR occurs when long, short, and local skip connections are all included.Performance decreases when any connection type is absent.
- Skip Connections: Increasing network depth without skip connections does not improve BSD68 performance.This observation contrasts depth alone with the contribution of skip connections.
- Feature Attention: Networks with feature attention achieve higher PSNR than corresponding networks without feature attention.The comparison supports the reported benefit of feature attention in the ablation study.
- Design Interpretation: The authors attribute further CNN denoising improvement to careful block design and feature-map rescaling, represented by feature attention and skip connections.These two characteristics are identified as design elements of RIDNet.
4.3. Comparisons
RIDNet is evaluated against traditional and CNN-based denoisers on synthetic grayscale and color data plus four real-noise datasets. It reports strong quantitative and qualitative performance, while RNI15 permits qualitative comparison only because clean images are unavailable.
- Test datasets: The evaluation covers four real-world datasets and three synthetic datasets, with synthetic images corrupted by additive white Gaussian noise at σ=15, 25, and 50.RNI15, DND, Nam, and SSID are used for real noise; 12 classical images and BSD68 grayscale and color images are used for synthetic noise.
- Evaluation constraints: RNI15 is evaluated qualitatively because its clean images are unavailable, whereas DnD reference images are not publicly available and scores come through an online system.These dataset constraints limit direct quantitative evaluation for RNI15 and local access to DnD ground truth.
- Grayscale noisy images: RIDNet outperforms all compared methods on Set12 and BSD68 grayscale images across all tested noise levels.The comparison includes nonlocal self-similarity, learning-based, and CNN-based denoisers, using PSNR.
- Color noisy images: RIDNet consistently achieves the best PSNR and perceptual quality on the CBSD68 color dataset.The visual comparison reports textures closest to ground truth, with fewer artifacts and more details.
- Real-world noisy images: On DnD, RIDNet leads CBDNet by 1.17dB PSNR and exceeds Noise Clinic and Neat Image by 3.82dB and 4.14dB, respectively.The DnD evaluation reports PSNR and SSIM from the online benchmark and compares against blind and non-blind denoisers.
- Real-world noisy images: On SSID, RIDNet exceeds FFDNet by 9.5dB and CBDNet by 7.93dB PSNR, while Nam results are higher than those of the compared state-of-the-art algorithms.Qualitative examples additionally report preserved structures and textures on DnD and RNI15, and recovered colors on SSID.
5. Conclusion
The paper presents a CNN denoising model for synthetic noise and real noisy photographs, using feature attention and low-frequency bypass connections. Experiments on three synthetic and four real-noise datasets demonstrate its effectiveness.
- The proposed CNN model addresses both synthetic noise and real noisy photographs.
- Feature attention rescales channel-wise features using dependencies between channels.
- LSC, SSC, and SC bypass low-frequency information so the network can focus on residual learning.
- Experiments on three synthetic and four real-noise datasets demonstrate the effectiveness of the proposed model.