Improved Automating Seismic Facies Analysis Using Deep Dilated Attention Autoencoders

Zengyan Wang, Fangyu Li, Thiab R. Taha, Hamid R. Arabnia; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 0-0

With the dramatic growth and complexity of seismic data, manual annotation of seismic facies has become a significant challenge. The encoder-decoder neural network architecture has been widely used in image segmentation. In recent years, the same architecture has also been used in seismic surveys for facies classification applications. In this paper, a modified U-Net architecture with trainable soft attention mechanism and dilated convolution is proposed to improve the automatic seismic facies analysis. This proposed framework generates more accurate results in a more efficient way. The dilated convolution achieves more accurate results with less computation than the CNN with pooling in U-Net. With the attention mechanism, the dilated U-Net model further improves classification accuracy. Our experiments show that the dilated attention autoencoder model is less prone to overfitting and at the same time, it achieves a smoother increasing validation accuracy.