Seismic interpolation based on quadratic denoising neural network

Seismic interpolation is crucial for seismic data processing by interleaving missing traces with synthesized traces to increase spatial sampling. Recently, a deep learning-based method is developed to achieve seismic interpolation by integrating the well-trained denoising convolutional neural network (DnCNN) into the project onto convex set (POCS) algorithm. It is flexible and applicable for different types of missing traces. However, the effectiveness of POCS algorithm is related to the denoising performance. To enhance reconstruction performance, we propose integrating a quadratic denoising neural network into POCS algorithm, leveraging its superior representation capability and robust denoising performance. The conventional convolutional layers are replaced by quadratic neurons, exploiting better nonlinearity. Numerical experiments illustrate that the proposed method provides better interpolation results. Moreover, the proposed approach reduces more aliases and leakages in frequency spectrum.