天然气工业2025，Vol.45Issue(11)：57-68,12.DOI:10.3787/j.issn.1000-0976.2025.11.005
基于深度多网络特征融合的测井成像图孔洞智能识别方法
An intelligent identification method for pores and vugs in logging images based on deep multi-network feature fusion
摘要
Abstract
The precise segmentation and identification of geological structures such as pores,vugs and fractures in logging images are crucial for the evaluation of oil and gas reservoirs.Although deep learning techniques have made some progress in this field,challenges still remain,including high false negative rates for small targets,strong background interference,and insufficient fusion of multi-scale features.To address these challenges,this paper proposes a deep multi-network feature fusion and segmentation model based on the synergistic collaboration of Convolutional Neural Network(CNN)and Transformer dual branches,and establishes an encoding network with CNN as the backbone and Transformer as the auxiliary branch.Then,a Partitioned Merge Channel Attention(PMCA)mechanism is designed in the decoder,which can effectively restore detailed information by performing spatial partition,weight learning and feature enhancement on feature maps.Finally,the model performance is validated using FMI logging datasets from multiple blocks,accompanied by systematic ablation experiments and comparative analysis.The following results are obtained.First,the dual-branch CNN-Transformer architecture enables cross-stage feature interaction,significantly enhancing the model's ability to characterize multi-scale pores and vugs.Second,the PMCA module integrated in the decoder effectively calibrates feature channel weights,achieving targeted enhancement of key features while suppressing background interference effectively.Third,the proposed model demonstrates an outstanding capability of capturing contours and edge features of pores and vugs in complex geological settings,with an overall segmentation accuracy and robustness surpassing current mainstream methods.In conclusion,on real datasets,the newly established model achieves 84.49％mIoU,89.76％precision,and 92.21％recall,significantly outperforming mainstream segmentation networks.What's more,Image Similarity(IS)and Accurate Segmentation Accuracy(ASA)are introduced as quantitative evaluation indexes,which reach 81.6％and 83.2％on the test set,respectively,demonstrating the model's potential to replace manual interpretation in the tasks of pore and vug segmentation and annotation.Furthermore,the new method provides efficient and accurate technical support for the automatic interpretation of logging images,and is of significant engineering guidance value for assessing wellbore and wall stability and quantitatively analyzing reservoir spaces in oil and gas resource exploration.关键词
测井成像图/深度多网络/孔洞分割/语义分割/Transformer/卷积神经网络/人工智能技术Key words
Logging image/Deep multi-network/Pore and vug segmentation/Semantic segmentation/Transformer/Convolutional neural network(CNN)/Artificial intelligence(AI)分类
天文与地球科学引用本文复制引用
夏文鹤,刘茗月,官文婷,丁星,陈向东..基于深度多网络特征融合的测井成像图孔洞智能识别方法[J].天然气工业,2025,45(11):57-68,12.基金项目
国家重点研发计划项目"超快响应湿度传感器在油气勘探、工业监控等领域的应用验证"(编号:2023YFB3210205). （编号:2023YFB3210205）
