应用MMTONet 的迁移学习智能盐体分割方法

打开文本图片集

关键词： 深度学习，盐体分割，地震图像，迁移学习，MMTONet 方法中图分类号：P631 文献标识码：A DOI：10. 13810/j. cnki. issn. 1000⁃7210. 20240270

Transfer learning⁃based intelligent salt body segmentation method using MMTONet

LI Kewen1，FAN Yating1，XU Zhifeng1，JIA Shaohui2 （1. Qingdao Institute of Software and College of Computer Science and Technology，China University of Petroleum（East China），Qingdao，Shandong 266580，China；2. Science and Technology Research Institute Branch，National Petroleum and Natural Gas Pipeline Network Group Co. ，Ltd. ，Langfang，Hebei ，China）

Abstract ：Salt bodies are geological structures with good airtightness ，which are conducive to oil and gas sto⁃ rage. It is extremely necessary to achieve refined interpretation of salt bodies. However，unlike faults，salt bo⁃ dies have more complex characteristics and significant morphological differences，and thus conventional meth⁃ ods can easily lead to confusion and misjudgment. In addition，since data ⁃ driven salt body recognition models have poor generalization ability on actual datasets，there are still challenges in interpreting and visualizing salt bodies in seismic exploration. The paper regards salt body interpretation as a semantic segmentation problem for seismic images and proposes an intelligent salt body segmentation method based on the context fusion of transfer learning and mixed attention （multi ⁃ path structure mixed attention and transfer optimized net， MMTONet）. At the same time，a salt body context feature fusion module is designed，and an improved atten⁃ tion convolution mixed skip connection mechanism is established to better compensate for the information loss caused by down⁃sampling ，thereby improving the pixel⁃level discrimination ability of the model for salt body boundaries and highamplitude noise. On this basis，a transfer learning adapter fine ⁃ tuning strategy is also de⁃ signed to improve the generalization ability of the model on actual data. The experimental results on seismic da⁃ tasets show that MMTONet outperforms mainstream semantic segmentation methods in improving segmenta⁃ tion accuracy and reducing computational and parameter complexity.

Keywords：deep learning，salt body segmentation，seismic image，transfer learning，MMTONet methodusing MMTONet［J］. Oil Geophysical Prospecting，2025，60（3）：631‑641.

0 引言

盐体是重要的储层结构，包含了有关油气的资源信息[1]。（剩余14716字）