Self-Supervised Image Denoising with Subsampling and Residual Learning

Image denoising based on deep learning has made more extensive development by using a large amount of data for network training, however, it is difficult to obtain clean images without noise in actual scenarios, leading to the emergence of self-supervised deep learning technique. In this step, we raise two questions: how to improve the performance of self-supervised learning and how adaptive is the self-supervised method to various networks. In response to these issues, we propose a self-supervised image denoising scheme which generates image training pairs by subsampling the noisy image twice, and combines U-Net and ResNet to form an effective denoising network. As a result, we improve the denoising performance of the latest self-supervised method Neighbor2Neighbor by building a deeper convolutional network while avoiding gradient explosion through careful setting of training parameters. The experimental results show that the proposed method improves the PSNR value by 0.3 dB and SSIM value by 0.004 on average compared to the original algorithm, and also verifies the robustness of the training strategy of Neighbor2Neighbor algorithm to different network structures.