Seismic Image Difference Analysis Based On Siamese CNN

In recent years,the data-driven artificial intelligence technology represented by neural network has developed rapidly in the field of seismic exploration such as the processing and interpretation of seismic data.ways and ideas.Among many deep neural networks,the Siamese convolutional neural network is a flexible infrastructure network that is good at processing paired images and other data,and has a wide range of applications in face recognition,change monitoring and other fields.This thesis is dedicated to the differential analysis and solutions of pre-stack seismic sections,and deeply studies the Siamese CNN and its application in the field of seismic exploration.The specific research contents are as follows:First,the Siamese CNN architecture is deeply studied,and a new Siamese CNN architecture is proposed,which includes the following innovations: The performance of feature maps in each stage is analyzed,and a hierarchical feature fusion module is proposed in the image feature fusion stage;skip connections of feature backhaul are added to feature maps at different depths;the upsampling stage of feature maps is integrated into the Siamese CNN,using dual-supervised combined training of two-stage loss functions to achieve end-to-end scene change detection.The new Siamese CNN improves the data utilization efficiency and prediction accuracy,laying a foundation for the subsequent application of the Siamese CNN to the seismic data processing stage.Secondly,in view of the problem of error differences in seismic data under complex geological structures at different offsets,an unsupervised seismic data registration method based on the Siamese CNN is designed in this thesis.The profiles are registered with the seismic profiles at near offsets through the coordinate transformation between the values,and the deformation fields are calculated for the seismic profiles at different offsets.Coordinate transformation for subsequent horizontal stacking of seismic sections at near offsets.The method is applied to the processing of the Sigsbee2 B seismic model data before the horizontal stacking after dynamic correction,and the results confirm that the method can greatly improve the horizontal stacking effect.