[Paper Review] DCANet: Dual Convolutional Neural Network with Attention for Image Blind Denoising

Wencong Wu, Guannan Lv+4 more|arXiv (Cornell University)|Apr 4, 2023

Image and Signal Denoising Methods4 citations

TL;DR

This paper proposes DCANet, a novel dual convolutional neural network with a spatial and channel attention module for image blind denoising. It integrates a noise estimation network, a dual-branch CNN for complementary feature learning, and an attention mechanism to suppress real-world noise effectively, achieving state-of-the-art performance with a favorable parameter-to-performance trade-off.

ABSTRACT

Noise removal of images is an essential preprocessing procedure for many computer vision tasks. Currently, many denoising models based on deep neural networks can perform well in removing the noise with known distributions (i.e. the additive Gaussian white noise). However eliminating real noise is still a very challenging task, since real-world noise often does not simply follow one single type of distribution, and the noise may spatially vary. In this paper, we present a new dual convolutional neural network (CNN) with attention for image blind denoising, named as the DCANet. To the best of our knowledge, the proposed DCANet is the first work that integrates both the dual CNN and attention mechanism for image denoising. The DCANet is composed of a noise estimation network, a spatial and channel attention module (SCAM), and a CNN with a dual structure. The noise estimation network is utilized to estimate the spatial distribution and the noise level in an image. The noisy image and its estimated noise are combined as the input of the SCAM, and a dual CNN contains two different branches is designed to learn the complementary features to obtain the denoised image. The experimental results have verified that the proposed DCANet can suppress both synthetic and real noise effectively. The code of DCANet is available at https://github.com/WenCongWu/DCANet.

Motivation & Objective

• Address the challenge of real-world image denoising where noise distribution and level are unknown and spatially varying.
• Overcome limitations of existing DNN-based models that are trained for specific noise types or levels, making them inflexible in practical scenarios.
• Integrate dual CNN branches and a spatial-channel attention mechanism to enhance feature learning and noise suppression.
• Achieve a balance between model complexity and denoising performance for practical deployment.
• Develop a fully blind denoising model that estimates noise distribution and level end-to-end without manual tuning.

Proposed method

• Employ a noise estimation network to predict spatial noise distribution and noise level from the input noisy image.
• Design a spatial and channel attention module (SCAM) that processes the noisy image and estimated noise to highlight relevant features and suppress irrelevant ones.
• Implement a dual-convolutional denoising network with two distinct branches to learn complementary features from the same input, enhancing representation capacity.
• Feed the noisy image and estimated noise into the SCAM, which then guides the dual CNN to reconstruct the clean image.
• Train the entire network end-to-end using a perceptual loss and L1 loss to preserve structural and texture details.
• Utilize skip connections and residual learning to stabilize training and improve feature propagation.

Experimental results

Research questions

• RQ1 Can a dual-branch CNN with an attention mechanism improve blind image denoising performance compared to single-branch models?
• RQ2 Does integrating a noise estimation network with attention and dual-branch learning lead to better generalization on real-world noise?
• RQ3 How does DCANet compare to state-of-the-art models in terms of PSNR, SSIM, and FLOPs on synthetic and real noise?
• RQ4 Can the proposed architecture achieve a favorable trade-off between model complexity and denoising accuracy?
• RQ5 Is the attention mechanism effective in filtering out noise while preserving image details in blind denoising settings?

Key findings

• DCANet achieves competitive PSNR and SSIM scores on both synthetic and real-world noise datasets, outperforming several SOTA models including FFDNet, BUIFD, and ADNet.
• On the real noise benchmark (NTIRE2020), DCANet achieves a PSNR of 0.258 dB on color images and 0.252 dB on grayscale images, demonstrating strong robustness.
• Despite having only 1.389 million parameters, DCANet achieves better or comparable performance than models with significantly more parameters, such as AirNet (8.9K parameters) and VDN (7.8K parameters).
• The ablation study confirms that the dual-branch structure and attention mechanism each contribute significantly to performance, with the combination yielding the best results.
• The model generalizes well across different noise levels and types, including spatially varying noise, due to the end-to-end noise estimation and adaptive feature learning.
• The FLOPs of DCANet (37.88 G) are moderate, indicating efficient inference, and the model maintains high performance with relatively low computational cost.