MMPDNet: Multi-Stage & Multi-Attention Progressive Image Denoising
Jiangbo Xue, Jiu Liang, Jinhe He, Yu Zhang, Yanda Hu · 2021
Deep convolutional neural networks (CNNs) have gained great success in many low-level tasks of computer vision, especially in the task of image denoising. However, there are two major drawbacks for denoising tasks, one is that it is very hard to choose a strategy to make deeper CNN convergence, other is that mainstream approaches do not do well a complex balance between spatial details and high-level features information while reconstructing clean images. In this paper, we proposed a novel multi-stage and multi-attention architecture of CNN for image denoising. In detail, our model is a multi-stage network, in each stage, we separate degraded inputs into different patches to learn the reconstruct mapping, thus, the overall denoising process is divided into several easily convergence steps. Meanwhile, a channel-spatial attention module is proposed to learn high-level features from different dimensions. Furthermore, an effectively supervised module is employed between every two-stage to refine the reconstructed results. The resulting multi-stage and multi-attention architecture, named MMPDNet, and its extensive experiment results deliver strong performance gains on some primary real-word denoising datasets, including SIDD and DND.