Denoising prestack random noise with deep generative prior

Supervised methods based on deep learning require noise-free training labels, either realistic-looking synthetic data or denoised results via conventional methods. Without ground-truth labels, it is challenging to get a considerable improvement in denoising effectiveness. Besides, generalization performance is also a critical problem for supervised methods. We proposed an unsupervised method based on a deep generator network to suppress random noise, which is a reconstruction process of prestack seismic data. Specifically, deep generator networks tend to generate signals with high correlation, while we pre-train the network to learn the mapping from random latent vectors to useful signals, limited by its own structure and pre-training, the output of the network is limited to the manifold of valuable signals. Then, we optimize the network parameters to find a point in this manifold closest to the noisy data. Such a point corresponds to the valuable signals contained in noisy data. To avoid the model overfitting the noise, both prior regularization and network structure regularization are used in our work. The random noise attenuation results of a CRP gather prove our method has an efficient denoising ability and a good generalization performance.