Research Advanced in Image Denoising Based on Deep Learning

Abstract

Image denoising has always been a hot research issue in the computer vision community, which aims to reduce the noise in digital images to improve image quality. As a key step in many downstream computer vision tasks, image denoising has been widely used in many fields such as medical agricultural images, satellite images, remote sensing images, face recognition, vehicle detection and many other fields. However, limited by problems such as insufficient actual image data sets, uneven noise reduction ability of different models, uneven noise reduction performance of the same model on different data sets and uneven noise reduction results of the same model for different scales of noise, the existing denoising algorithms still can not fully meet the needs of practical applications. In order to explore the optimal denoising method under the influence of the above variables, we first introduce three representative denoising frameworks based on self-supervised, residual network and generative countermeasure network. We then choose five best noise reduction models: CBM3D, DnCNN, IRCNN, FFDNet, and Fdncnn, and classify and compare the real noise and synthetic noise on five data sets, such as CBSD68, Kodak24, McMaster, DND, PolyU. In addition, in the noise reduction experiment of synthetic noise, we divide the noise into five noise levels, and discuss the change trend of noise reduction effect when the noise level changes. This paper has done a lot of experiments, but it still has high scalability. In the future, more independent variables can be introduced on the basis of this experiment, so as to draw more accurate conclusions.