Random Noise Attenuation Using Deep Convolutional Autoencoder

Abstract

Summary Suppressing random noise is very important to improve the signal-to-noise ratio of seismic data. We propose a novel method to attenuate random noise using deep convolutional autoencoder, which belongs to the unsupervised feature learning. We directly use the noisy data rather than a relatively noise-free data as the training target to construct the cost function and design a robust convolutional autoencoder network that can achieve random noise attenuation. Therefore, we always have an available input dataset to train the neural network, which can save us the trouble of seeking a relatively clean data. We use normalization and patch sampling to build training dataset and test dataset from raw seismic data. The back-propagation algorithm is used to optimize the cost function. The optimized parameters of convolution filters can be obtained after a stable optimization. The final denoised result can be reconstructed via the optimized convolutional autoencoder. Real data test proves the effectiveness of the proposed method.