Seismic Image Denoising with Zero Shot Strategy

Abstract We propose a data-free seismic denoising network based on the ideas of Zero-Shot Noise2Noise from pioneering deep learning techniques for natural image processing. The idea involves generating a pair of noisy images from a single noisy image and subsequently training a small network on this pair. The method engages the use of a downsampling operator to decompose an image into a pair of downsampled images. A lightweight network is subsequently trained using regularization techniques to map one downsampled image to another. The loss function consists of two terms: residual loss and consistency loss. The residual loss in denoising neural networks is formulated using only the downsampled images as input, while the network receives the image in full spatial resolution for the consistency loss.The key insight of this network is that the network can learn to denoise seismic images without any training data or knowledge of the noise distribution. This is achieved by training the network in a self-supervised manner. Numerical experiments demonstrate the validity of the network on seismic data.