OGFNet: Original Resolution Subnetwork and GRU Based Feature Fusion Network for Image Denoising

Abstract

Many denoising models have a good performance on images containing spatially invariant noise, especially those based on Convolutional Neural Networks (CNN). Nevertheless, the result of these methods is unreliable in preserving spatial image details. In our paper, we put forward a new design that achieves the goal of image denoising as well as preserve image details. In recent years, attention mechanisms is more and more popular in computer vision. We have done some researches on attention mechanisms in image denoising. Firstly, a combination of residual network and channel attention mechanisms and Gated Recurring Unit (GRU) is proposed to boost the denoising performance. Significantly, the GRU extracts features from different layers and selects more refined image features. Secondly, Original Resolution Subnetwork (ORSNet) is proposed to preserve the structural and textural image details. Few methods use a subnetwork for detail preservation. Extensive experiments on public denoising datasets attest to the validity of our proposed model.