Spatially adaptive EPLL denoising for low-frequency seismic random noise

The expected patch log likelihood (EPLL) frame utilizes a Gaussian mixture model (GMM) learned from external data as signal priors. The EPLL denoises the image patches via their most likely Gaussian component in the GMM and weighted-average the denoised patches and the noisy image to reconstruct the denoised image, leading to asuccessful denoising performance for random noise in the seismic image. Since the regularization parameter is only associated with the noise variance, it is difficult to achieve the balance between weak signal preservation and noise suppression for the desert seismic images containing non-strationary seismic signals and low-frequency colored noise. However, the EPLL is unadaptable to the non-stationary seismic signals in desert seismic images. A spatially adaptive EPLL (SA-EPLL) algorithm is proposed in this paper under the framework of the EPLL. In this method, we stabilize the seismic image with the variance normalization method and construct the patch signal-to-noise ratio (P-SNR) related regularization parameter, so that it can be adaptively adjusted with the spatiotemporally various intensity of the non-stationary seismic signals, allowing the balance between the preservation of local details and the restoration of global features of the non-stationary signals. In addition, in the signal reconstruction process, the P-SNR is used as the weight to weighted-average the denoised image patches, leading to a better denoising performance in less signal loss. The SA-EPLL algorithm is applied to synthetic and field seismic images, with the results showing that the proposed method can effectively restore non-stationary signals and suppress low-frequency random noise with weak similarity in desert seismic images. © 2021, The Editorial Board of Journal of Xidian University. All right reserved.