Multiscale Dense Attention Network for Seismic Data Reconstruction

TL;DR: Wang et al. as discussed by the authors proposed a novel multiscale network structure combined with dense spatial attention mechanism to further promote the performance of CNN-based reconstruction methods for seismic data, which contains three scales which can extract multi-scale features from seismic data with different resolutions.

Abstract: Summary Reconstruction is one of the essential steps in the seismic data processing. Effective reconstruction methods can not only provide high-density data volume for the following high-resolution seismic imaging, but also alleviates the data acquisition cost in seismic exploration. Recently, deep-learning-based methods, especially convolution neural networks (CNN), are gradually applied to the reconstruction of seismic data and have shown remarkable reconstruction performance. However, most of these CNN-based reconstruction methods only consider features in single resolution or just utilize simple interactions between different scales, which is likely to result in performance degradation when dealing with some complex seismic data. To further promote the performance of CNN-based reconstruction methods for seismic data, we propose a novel multiscale network structure combined with dense spatial attention mechanism. Specifically, this proposed network contains three scales which can extract multi-scale features from seismic data with different resolutions. Moreover, we utilize the dense spatial attention module to fuse these multi-scale features. Experimental results demonstrate that this proposed multi-scale network exhibits better performance in seismic data reconstruction compared with two conventional methods and the classical U-Net.