Fast Dictionary Learning With Automatic Atom Classification for Seismic Data Denoising

Seismic noise attenuation is a long-standing yet still challenging topic in seismic data processing. Dictionary learning (DL) has emerged as an effective way to attenuate spatially incoherent noise without damaging useful signals. DL can adaptively learn atoms that best represent the structure of seismic data in a fully data-driven way, and thus can be applied in arbitrarily complex seismic datasets. Due to the existence of strong noise, the dictionary atoms could be strongly affected, e.g., containing many atoms that show quite irregular structures. Here, we propose a new method for effectively rejecting those noise-contaminated atoms while maintaining the most representative atoms, so as to improve the denoising performance. Atom classification is performed by taking advantage of the statistical features of the atoms in each category, e.g., regular (for signal) and irregular (for noise). More importantly, the rejection of atoms with a high probability of noise inference is controlled easily by a threshold, below which the atom is preserved for sparse coding and denoising. An efficient dictionary updating scheme is also used to avoid the computationally expensive singular value decomposition (SVD). We demonstrate the performance of the proposed method in both synthetic and real datasets.