DMANet: An Image Denoising Network Based on Dual Convolutional Neural Networks with Multiple Attention Mechanisms

Addressing the limitations of current convolutional neural network-based image denoising methods in capturing local features and reconstructing global details, as well as the potential issues of vanishing or exploding gradients caused by stacking multiple convolutional layers, the paper proposes an image denoising network based on dual convolutional neural networks and multi-attention mechanisms, named as the DMANet. To tackle the problem of multi-scale feature extraction, a Multi-branch Feature Extraction Module is presented. This module consists of global, large-scale, and local branches, comprehensively capturing multi-scale features from global to local scales of the image. A Supervised Attention Module is incorporated between the traditional dual convolutional neural networks, which can filter out and suppress unimportant feature information during the information transmission process, thereby improving the efficiency of network information extraction. To reduce model complexity, the spatial and channel attention module in DCANet is optimized by replacing the original Channel Attention (CA) mechanism with a simplified CA. Experiment results demonstrate the robustness and effectiveness of DMANet under various noise conditions, and enhance the denoising effect.