Two-stage Progressive Residual Dense Attention Network for Image Denoising
Wencong Wu, An Ge, Guannan Lv, Yuelong Xia, Yungang Zhang, Wen Xiong
TL;DR
The paper addresses image denoising under additive noise by recovering the clean image $x$ from a noisy observation $y$ with $x = y - N$, including real-world noise. It introduces a two-stage progressive network, TSP-RDANet, composed of a residual-dense attention stage (RDAM) in Stage 1 and a hybrid-dilated residual-dense attention stage (HDRDAM) in Stage 2, with long skip connections to fuse shallow features. RDAM/HDRDAM leverage dense connections and attention to emphasize informative local features while suppressing irrelevant ones, and employ downsampling and dilated convolutions to enlarge the receptive field. Experiments across seven datasets demonstrate competitive or state-of-the-art denoising performance for both synthetic and real noise, with favorable speed/parameter trade-offs; code is available at the provided GitHub link.
Abstract
Deep convolutional neural networks (CNNs) for image denoising can effectively exploit rich hierarchical features and have achieved great success. However, many deep CNN-based denoising models equally utilize the hierarchical features of noisy images without paying attention to the more important and useful features, leading to relatively low performance. To address the issue, we design a new Two-stage Progressive Residual Dense Attention Network (TSP-RDANet) for image denoising, which divides the whole process of denoising into two sub-tasks to remove noise progressively. Two different attention mechanism-based denoising networks are designed for the two sequential sub-tasks: the residual dense attention module (RDAM) is designed for the first stage, and the hybrid dilated residual dense attention module (HDRDAM) is proposed for the second stage. The proposed attention modules are able to learn appropriate local features through dense connection between different convolutional layers, and the irrelevant features can also be suppressed. The two sub-networks are then connected by a long skip connection to retain the shallow feature to enhance the denoising performance. The experiments on seven benchmark datasets have verified that compared with many state-of-the-art methods, the proposed TSP-RDANet can obtain favorable results both on synthetic and real noisy image denoising. The code of our TSP-RDANet is available at https://github.com/WenCongWu/TSP-RDANet.