Fast convex set projection with deep prior for seismic interpolation

Reconstruction of missing traces from seismic data is traditionally handled by physical methods with good interpretability. Popular deep learning methods provide promising end-to-end solutions, however requiring a large amount of labeled data. Currently, there is a trend to fuse physical and deep learning methods to take advantage of both. In this paper, we propose to incorporate the convolutional neural network denoising model for image restoration (IRCNN) into the physical fast convex set projection (FPOCS) framework for seismic data interpolation. IRCNN is an off-the-shelf denoising model that has been pre-trained with abundant natural images. We fine-tune it with a few thousands of synthetic and field seismic patches. Consequently, the new algorithm, which will be called IRCNN-FPOCS, has three advantages: (1) supports high-performance seismic interpolation with deep priors; (2) is interpretable; and 3) alleviates the problem of insufficient training data for the seismic field. Experiments are conducted on regularly and irregularly subsampled synthetic and field data in comparison with the Monte Carlo data-driven tight framework (DDTF) and the convex set projection method with CNN prior (CNN-POCS). Results show that our method is superior to the counterparts in terms of (1) visual effects and quantitative evaluation indicators; and (2) generalization ability to seismic data with different sampling ratios.

Automated summary

This paper proposes a new algorithm that combines deep learning and physical methods for seismic data interpolation. Experimental results show that the proposed method outperforms other methods in terms of visual effects and quantitative evaluation indicators, and it also has interpretability and generalization ability to different seismic data.