Leveraging overfitting for good - Two-Step Deep Image Prior Model For Seismic Denoising

Accurate separation of signal and noise constitutes a fundamental prerequisite for achieving high-resolution seismic imaging. A notable recent advancement in this domain is the Deep Image Prior (DIP), an unsupervised deep learning method leveraging deep neural networks (DNNs). The success of this approach lies in the adoption of autoencoders that enables the adaptive extraction of high-fidelity data features. However, establishing an optimal balance between noise suppression and signal preservation remains a non-trivial challenge for DIP-based seismic denoising methods, which is affected by the potential issue of overfitting. This arises from the inappropriate selection of a network architecture and the corresponding hyperparameters, especially the number of training epochs, which strongly influence the learning capacity and feature extraction capabilities of the model. In response to this challenge, we introduce Two-Step DIP (TSDIP), a novel denoising method that exploits overfitting to enhance seismic data quality. In the initial stage, the proposed DNNs are intentionally trained to overfit by effectively attenuating high-frequency noise from the input data. Subsequently, the proposed DNNs are employed iteratively to suppress any residual noise in the newly processed data without damaging useful signal. The overfitting in the first step helps precondition the data to be at a lower noise level while preserving as much as fine-scale features in the signal. To employ an optimal balance, we carefully determine an ideal number of epochs, which is consistently applied in both denoising steps. To assess the effectiveness of the TSDIP method, we present the test results derived from 3D synthetic and field seismic datasets. Our analysis indicates that TSDIP effectively reduces strong noise while preserving key seismic details through the use of overfitting.