Drop-DIP: A single-image denoising method based on deep image prior

Abstract

. Over the past few years, deep learning methods have emerged as powerful image denoising tools. Among them, unsupervised deep learning without external training data is more practical and challenging. Reducing noisy overfitting is challenging due to single-image unsupervised learning is prone to overfitting. In this paper, we propose a method named drop-DIP combing Deep Image Prior (DIP) with drop-out for the first time to solve the above problems. In our method, we construct new network training pairs by performing drop-out training on the Bernoulli sampling of the input and output, and then construct a regularization term by using the corrected bias of the output and the generated prior. Finally, update the parameters through the Alternating Direction Method of Multipliers (ADMM) algorithm. Experiments demonstrate that drop-DIP can alleviate the overfitting difficulty in DIP, facilitate the early stopping of the network, and is applicable to different noise models. Furthermore, our method has good performance on Peak Signal to Noise Ratio (PSNR), Structural Similarity (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) metrics validated by two different datasets.