Can Knowledge Distillation Enable Seismic Interpolation, Super-Resolution, and Denoising Simultaneously?

In seismic data processing, interpolation, super-resolution, and denoising are three key issues to improve data quality. Various deep learning methods have gained success in handling one or two of them. However, addressing all three with a single network faces difficulties such as conflicting goals and complex network design. This paper proposes a heterogeneous knowledge distillation architecture with a teacher network (TNet) and a student network (SNet) to handle this challenging task. TNet is pre-trained with noisy low-resolution data, while SNet is trained with missing noisy low-resolution data, avoiding conflicting goals for both networks. SNet shares the same structure as TNet but with half the channels, making both networks simpler. SNet adopts a combination of data consistency loss and distillation loss, with the former being a combination of ℓ<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub> and multi-scale structural similarity, while the latter is synchronous positive negative learning. Therefore, it inherits the capabilities of TNet and further enhances them during the optimization process. For comparison, we combine three interpolation algorithms with three super-resolution algorithms and select the one with the best performance. Experimental results on four field data show the superiority of our method over the counterpart in terms of data quality at a lower computational cost.