Attention-Based 3-D Seismic Fault Segmentation Training by a Few 2-D Slice Labels

Detection faults in seismic data are a crucial step for structural interpretation, reservoir characterization, and well placement. Some recent works regard it as an image segmentation task. The task of requires huge labels, especially 3-D data, which has complex structure lots noise. Therefore, its annotation expert experience workload. In this study, we presented $\lambda $ -binary cross-entropy (BCE) -smooth notation="LaTeX">$L_{1}$ loss to effectively train 3D-CNN by some slices from volume label, so that the model can learn few 2-D slices. order fully extract information limited suppress noise, proposed attention module be used active supervision training embedded network. map label is generated original letting supervise using loss. experimental results demonstrate function features slice labels. And also shows advanced performance module, significantly noise while increasing sensitivity foreground. Finally, on public test set, method achieved similar labels only 3.3%