Research on Image Denoising Based on Dual-Path Multi-scale Feature Fusion Network

With the development of deep learning, images make significant progress in denoising. In order to make the denoising effect better, this paper proposed a dual-path multi-scale feature fusion denoising network. The network structure adopts a dual-path mode, the difference between the two links are increased by using dilated convolution in one of the links, which is treated as a special downsampling operation. Compared with the link without dilated convolution, it can capture image features of different scales by expanding the receptive field. A multi-scale feature fusion module Connection Block (CB) is added to the network, which can combine image feature information at different levels to better preserve the texture and details of the image. In the residual module, every two convolution blocks are combined into a residual block which enables the network to achieve a better learning effect. Experiments demonstrated that using dual-path structure not only improves the denoising effect, but also expands the width of the network. Meanwhile, the network proposed in this paper mostly outperforms several state-of-the-art deep learning methods on average for removing grayscale and color images.