Towards Using Neural Networks to Complement Conventional Seismic Processing Algorithms

Convolutional-based neural network (CNN-based) architectures have shown promise in performing denoising tasks. However, it can be demonstrated that their predictions are of limited use for some tasks because they produce signal leakage. For these tasks, a possible improvement is to incorporate CNN-based architectures as one component of, rather than replacement for, the conventional denoising algorithms. In this paper, we formally define a class of denoising problems usually solved iteratively for which using CNN-based predictions as an initial solution can improve efficiency. We illustrate our points using a land data deblending example, for which the CNN-based prediction quality was higher than that of the conventional first iteration but lower than that of the final product. The CNN-complemented conventional deblending leads to satisfactory and efficient results.