| With the gradual maturity of artificial intelligence technology and the increasingly prominent effectiveness in practice,a series of algorithms represented by deep learning are also being applied by many researchers to traditional geophysical problems such as seismic data processing.In addition,thanks to the development of modern storage technology and the huge improvement of computer computing power,under the support of huge historical seismic data,deep nonlinear neural networks are enough to fit complex high-dimensional functions,which also makes the prediction results of the algorithm as close as possible to the real label results of the data.Different from the reconstruction of three-dimensional holes or fissures using geophysical modeling and derivation of mathematical formulas,this paper uses deep learning semantic segmentation algorithm to study the three-dimensional hole characterization problem,considering the need to process large-scale seismic data and the time spent in the model training stage,the newly constructed model will improve the recognition accuracy and inference speed at the same time.In this paper,the model is first built and improved for the three-dimensional salt body segmentation task,so that the model can achieve the ideal effect in accuracy and reasoning speed,and realize the hole characterization of the three-dimensional seismic data.After analyzing the applicability of a series of commonly used semantic segmentation models for salt segmentation tasks,the Deeplab V3 network with high accuracy and good plasticity is selected as the basic model for this task.Aiming at the defect that the Deeplab V3 network fails to make full use of the shallow features of the model,an auxiliary classifier that can make full use of the shallow features of the model is constructed,which greatly improves the accuracy of the model.In addition,by changing the feature extraction module of the network,the inference speed and accuracy of the model can be greatly changed,and the attention module can be added to the backbone network to stably improve the accuracy of the model.After the above improvements,the inference speed of the Deeplab V3 model with Mobile Net as the feature extraction module is only 21 ms,and the accuracy rate of the Deeplab V3 model with Res Net50 as the feature extraction module can reach 94.4%.Based on the above research,in order to make the accuracy of the model and the inference speed meet the expected effect at the same time,the knowledge distillation model is built by using the above two improved networks,and the Att＿Mobile Net model with faster reasoning speed but lower accuracy rate is used as the student model,The Att＿Res Net50 model with high accuracy but complex structure is used as the teacher model,and the shallow feature learning module and the deep feature learning module designed in this paper are used as the structure of information exchange between the two models to improve the accuracy of the student model,While improving the accuracy,it does not increase the number of parameters of the student model.This allows the final inference speed and accuracy of the student model to meet the expected requirements at the same time.The experimental results show that the trained knowledge distillation model can detect the punctate non-salt region and the narrow salt body region respectively.Finally,the trained knowledge distillation model is used to predict the three-dimensional seismic data body,and the hole characterization of the three-dimensional seismic data is realized. |
