Improving Performance of Seismic Fault Detection by Fine-Tuning the Convolutional Neural Network Pre-Trained with Synthetic Samples

Fault interpretation is an important part of seismic structural and reservoir characterization. In the conventional approach, faults are detected as reflection discontinuity or abruption manually tracked in post-stack data, which time-consuming. order to improve efficiency, a variety automatic fault detection methods have been proposed, among widespread attention has given deep learning-based methods. However, learning techniques require large amount marked samples training dataset. Although synthetic data can be guaranteed labels accurate, difference between real still exists. To overcome this drawback, we apply transfer strategy performance by We first pre-train neural network with data. Then retrain samples. use random sample consensus (RANSAC) method obtain generate corresponding automatically. Three 3D examples included demonstrate that accuracy pre-trained models greatly improved retraining few