Self2Align: a self-supervised denoising framework for single-scene images

Abstract

In recent years, driven by the development of deep neural networks, image denoising technology has made great progress. One of the most representative technologies is the deep neural network denoiser based on supervised learning. The denoiser uses noise-clean image pairs as the input and output of a deep neural network, and trains the deep neural network to achieve the denoising goal. However, collecting many high-quality noise-clean image pairs is extremely challenging, which is mainly because (1) it is difficult to collect true and clean images; and (2) changes in motion and lighting make it impossible to align collected image pairs, which limits the widespread application of supervised learning denoising techniques. To solve the above problems, this paper proposes a simple and effective method Self2Align to train a deep neural network denoiser. First, we proposed an efficient deep network model for inter-image alignment. For the collection of original images, we only use noisy images and collected multiple images for different scenes. The trained alignment network was then used to align the original image pairs automatically. Second, the aligned image pairs generated in the first stage were used as training image pairs for the training of the denoising network. In addition, we introduced a new training strategy so that the network can obtain better performance. The proposed Self2Align architecture eliminates the reliance on noise-clean image pairs and reduces the acquisition difficulty of training image pairs in terms of self-supervised training of the network. We explained the feasibility of our proposed method through theoretical analysis and obtained competitive results through experimental verification.