Boosting self-supervised blind-spot networks via transfer learning

Self-supervised procedures offer an appealing alternative to supervised denoising techniques that require noisy-clean pairs of training data. However, the capabilities of self-supervised denoising procedures are often limited by the requirement that noise cannot be predicted directly from neighbouring values in the training input samples. As such, there is often a trade-off with respect to the number of training epochs between learning to replicate the signal without learning to replicate the noise. Focusing on blind-spot networks that learn a pixel's value based on neighbouring pixels, we propose to train a supervised model in a blind-spot manner such that the model learns how to predict a pixel's clean value based off its noisy neighbouring traces. The weights of the trained model are then used to initialise a self-supervised model which is trained purely on noisy field data. In comparison to the fully self-supervised approach, we illustrate that pre-training with synthetic data results in increased noise suppression, alongside a lower level of signal leakage in the field data.