Contourlet Transform Based Seismic Signal Denoising via Multi-scale Information Distillation Network

Abstract

Recently, convolutional neural network (CNN) based models have achieved great success in many multimedia applications. Seismic signals as one kind of important multimedia resources are often interfered by noise in practice, which bring difficulties for utilizing these resources effectively. To this end, this paper presents a novel CNN architecture in the contourlet transform (CT) domain for seismic signal denoising. First of all, we propose multi-scale information distillation (MSID) module to fully exploit features from seismic signals. And then, a series of MSIDs are cascaded in a coarse-to-fine manner to restore the noisy seismic signals, which is named as deep multi-scale information distillation network (D-MSIDN). In addition, we propose to formulate the denoising problem as the prediction of CT coefficients, which is able to make D-MSIDN further remove noise and preserve richer structure details than that in spatial domain. We use synthetic seismic signals and public Society of Exploration Geophysicists (SEG) and European Association of Geoscientists & Engineers (EAGE) salt and overthrust seismic data to demonstrate the superior performance of the proposed method over other state-of-the-art methods. From some qualitative and quantitative results, we find that our method is capable of obtaining data with higher quality assessment and preserving much more useful signals than other methods. In particular, our denoising performances are more considerable for seismic data with lower signal-to-noise ratio.