A Self-Supervised CNN for Image Denoising with Self-Similarity Prior

Recently, a kind of blind-spot based self-supervised learning denoising method has attracted extensive research. The key of this kind of methods is to input Bernoulli-sampled instances and train network to recover the unsampled pixels. Based on the assumption that the image pixel is locally correlated while the noise exhibits statistical independence, the network will only recover the clean signal. Non-local self-similar priors play an important role in traditional image denoising methods, and can provide effective information for the reconstruction of unsampled pixels. We take blind-spot based method one step further by introducing non-local self-similarity prior into network processing flow. Specifically, we take the similar patch group as the processing unit, and design a non-local module in the network architecture to fuse the local and non-local information. Experiments show that the proposed non-local module can significantly improve the denoising performance.