A dual residual dense network for image denoising

Digital images have applications in almost every field, such as medicine, autonomous vehicles, astronomy, etc., for various purposes. The performance of all these applications depends on the quality of the images; that is, digital images should be noise-free for better results. Due to various factors, images capture noise at the acquisition time, which needs to be removed before subsequent applications. Therefore, this paper presents a dual residual dense network for image denoising. The network is designed using three types of blocks: feature extraction blocks, combination blocks, and residual blocks. These blocks exploit residual learning, dense connectivity, and batch renormalization to remove image noise. The network is also designed to be wide rather than deep, which makes it computationally and time-efficient. The network was trained on the DIV2K dataset and tested on the Kodak24 and Berkeley Segmentation (BSDS300) datasets. The results show that the proposed network outperforms existing state-of-the-art architectures in image denoising, achieving peak signal-to-noise ratio (PSNR) scores of 35.03 for the Kodak dataset and 34.64 for the BSDS300 dataset, as well as structural similarity index measure (SSIM) scores of 0.99 for both datasets. The code of the proposed method is publicly available to allow others to reproduce the findings and validate the results.