EDnNet: An Efficient U-Shaped Image Denoising Network

Abstract

Recently, a popular architecture, Transformer has garnered significant attention and success within the domain of natural language processing. However, due to its unique mechanism, the computational complexity increases exponentially with the spatial resolution of the input images. Therefore, it is not suitable for most tasks involving high-resolution image restoration. Although the receptive field of convolution is limited, the convolutional layer extracts local features by sliding small convolutional kernels on the input images. This ability allows Convolutional Neural Networks (CNNs) to capture essential local structures and texture information in images, which are crucial for image restoration. In our work, we propose a fundamental denoising block that relies on convolution, activation functions and normalization. Furthermore, we propose a denoising network named Efficient Denoising Network (EDnNet).