Real Image Denoising with Feature Attention (1904.07396v2)

Abstract: 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.

• The paper proposes a single-stage blind image denoising network (RIDNet) that achieves state-of-the-art performance.
• It introduces a feature attention mechanism to enhance key features and effectively manage spatially variant noise.
• Experiments on synthetic and real datasets show RIDNet outperforms 19 methods, including a 1.17 dB PSNR boost on the DnD dataset.

Real Image Denoising with Feature Attention

The paper "Real Image Denoising with Feature Attention" by Saeed Anwar and Nick Barnes introduces a novel approach for denoising images using a single-stage blind real image denoising network named RIDNet. The authors propose a method that improves on the limitations of existing denoising methods, particularly when dealing with real-world noisy images, which often entail spatially variant noise.

Key Contributions

The authors highlight several contributions demonstrating the efficacy and novelty of their approach:

1. Single-Stage Denoising: The proposed RIDNet achieves state-of-the-art performance using a single-stage model, while most other algorithms require multiple stages to handle the complexity of real image noise.
2. Feature Attention Mechanism: RIDNet incorporates a feature attention mechanism to selectively enhance significant features while suppressing less important ones. This focus on channel dependencies is a distinguishing characteristic of the proposed model.
3. Residual on the Residual Architecture: The network structure includes modular components that facilitate information flow, addressing vanishing gradient issues and improving scalability and performance as the network depth increases.
4. Comprehensive Evaluation: The authors evaluate RIDNet on both synthetic noisy datasets and real noisy datasets, demonstrating superior quantitative and qualitative results compared to 19 state-of-the-art algorithms.

Methodology

RIDNet's architecture consists of several notable elements, including a feature extraction module, a residual on the residual learning module, and a reconstruction module, which combine to effectively manage both synthetic and real-world image noise.

• Feature Extraction: This initial module captures essential features from the input noisy images using a convolutional layer.
• Feature Learning with Enhancement Attention Modules (EAMs): This component employs cascaded EAMs that include feature attention units and utilize globally pooled information to modulate channel-wise feature map weights.
• Modular Structure: Enhances performance through localized and long skip connections, which improve knowledge sharing across layers.

Experimental Results

The experimental results validate RIDNet's effectiveness across diverse datasets, including the DnD, RNI15, Nam, and SSID datasets. The method demonstrated superior performance in terms of PSNR and qualitative assessments, successfully addressing the challenges posed by real-image noise. Notably, RIDNet achieved a 1.17 dB improvement in PSNR over its closest competitor on the DnD dataset, showcasing its robustness and adaptability.

Implications and Future Directions

The approach presented in this paper enhances the practical applicability of image denoising techniques, emphasizing real-world conditions where models must perform under diverse and unpredictable noise conditions. The integration of feature attention with residual learning could inspire further research into leveraging attention mechanisms for other image restoration tasks.

Future developments may include exploring different pooling strategies for the attention mechanism or extending the modular architecture to more complex networks, broadening the application scope beyond image denoising to tasks such as super-resolution and image segmentation.

By balancing architectural innovations with practical performance requirements, this work contributes significantly to advancing the state of image denoising technology, providing a foundation for future research in real-world image restoration challenges.