Denoising method of borehole acoustic reflection image using convolutional neural network

Abstract

Using a small remote reflection wave, the borehole acoustic reflection imaging obtains high-resolution images from tens of meters away. Even after elimination, the direct mode wave excited near the borehole significantly contaminates the usable reflection wave, which then translates to solid noise in the migratory image. The severe noise significantly obstructs the application of the technique. The paper considers the use of a convolutional neural network for the primary data that we find most useful and treats the migration result as a type of digital image. We create a substantial dataset of simulated images using the Born approximation and the Kirchhoff migration method; we build a large-simulation image dataset. The designed multiscale recursive residual network leverages three branches to extract features at multiple resolutions. The learned multiscale features are recalibrated by the Double Attention Unit, and then aggregated by the Selective Kernel Feature Fusion module, preserving spatial details and extracting enriched contextual information. The network is trained on the simulated dataset using the MSE loss function and achieves higher accuracy in processing test images, compared with the widely-used DnCNN, demonstrating its effectiveness. The trained network is then applied directly to the real field image, where it successfully removes background noise while preserves the edges of valuable reflectors, proving the robustness and validity of the proposed methods.