In order to improve the efficiency and quality of salt-and-pepper denoising and realize the strategy of ‘denoising after judging’, the noise image classification network (CNN-J) is needed to judge whether the input image is a noisy image. For noisy images, the noise marking network (CNN-M) and noise denoising network (CNN-D) are combined for denoising processing, and the clean image will be directly output. In order to train the above three networks, three datasets are provided here, which are dataset_J, dataset_M and dataset_D, respectively.