Single Image Denoising through Downsampling and Self-Resolution Restoration Learning

Image denoising using supervised learning effectively removes image noise by learning from available data. However, it may lack efficiency when faced with insufficient data, such as in the case of single images or blind noise. This challenge has led to the adoption of unsupervised learning methods, which utilize the inherent properties of noise to extract and enhance image features. This research aims to leverage the benefits of the downsampling effect for noise removal, even though downsampling may impact image features. Therefore, deep learning must be used to restore image details lost during downsampling. This research proposes the Noisy Low-Resolution to Noisy Super-Resolution (NLR2NSR) framework, which leverages image downsampling to simultaneously reduce image and noise features. A super-resolution network is then used to restore the image features. Experimental results show that under conditions where noise features are less prominent than image features, the NLR2NSR can effectively remove noise and preserve image features using only noisy data for training. However, the NLR2NSR has limitations in handling high-level noise.