Deep nonlinear seismic prior for seismic interpolation

Seismic interpolation is an effective technology of reconstructing missing traces to improve the quality of seismic data. Over the past years, the deep learning methods show their powerful performance on seismic interpolation using a convolutional neural network. Recently, an unsupervised deep seismic prior method applies a convolutional neural network to generate the missing traces with one training sample. This method is convenient and suitable for regular and irregular missing seismic data. This method is convenient and suitable for regular and irregular missing seismic data. However, this neural network is based on the linear neurons, which suffer from its limited expressive ability for complex seismic signals. To enhance the capability of high-frequency information using stronger non-linearity neurons, we propose a deep nonlinear seismic prior method for seismic interpolation. Each convolutional layer is replaced by a nonlinear neuron, which is represented by a high-order polynomial of input data and weighted parameters. Numerical experiments illustrate that the proposed method provides better reconstruction results. Moreover, the proposed deep nonlinear seismic prior method learns the useful structure information faster and earlier than the conventional deep seismic prior method.