FR-UNet: A Feature Restoration-Based UNet for Seismic Data Consecutively Missing Trace Interpolation

Abstract

Convolutional neural network (CNN) is widely used for seismic data recovery and has demonstrated remarkable performance in reconstructing irregularly and regularly sampled seismic data. However, when it comes to recovering consecutively missing traces, CNN encounters difficulties in interpolating large gaps from the surrounding neighborhoods, due to the local property of convolution operator. Excessive missing entries existed in feature maps result in incomplete interpolation results. Thus, we propose a feature restoration-based UNet (FR-UNet) to improve the quality of reconstruction by restoring feature maps. In FR-UNet, we integrate feature recovering through the implementation of an attention transfer module (ATM). This module learns an attention score map from the high-level feature map of UNet, providing guidance for repairing the adjacent low-level feature map. Moreover, to ensure the integrity and precision of the highest level feature map, we utilize partial convolution (PConv) as a replacement for conventional convolution (CConv). Experimental results on synthetic and field data demonstrate that our network generates more accurate results for recovering large gaps through feature restoration.