基于迁移学习及通道先验注意力机制的地质构造识别  

Geological Structure Recognition Based on Transfer Learning and Channel Prior Attention Mechanism

作　　者：刘俊杰 马凯[1,2] 黄泽华[1,2] 田苗 邱芹军 陶留锋 谢忠 LIU Junjie;MA Kai;HUANG Zehua;TIAN Miao;QIU Qinjun;TAO Liufeng;XIE Zhong(Hubei Key Laboratory of Intelligent Visual Monitoring for Hydropower Engineering(China Three Gorges University),Yichang Hubei 443002,China;College of Computer and Information Technology,China Three Gorges University,Yichang Hubei 443002,China;Key Laboratory of Geological Survey and Evaluation of Ministry of Education(China University of Geosciences),Wuhan Hubei 430074,China;School of Computer Science,China University of Geosciences,Wuhan Hubei 430074,China;National and Local Joint Engineering Laboratory of Geographic Information Systems(China University of Geosciences),Wuhan Hubei 430074,China)

机构地区：[1]湖北省水电工程智能视觉监测重点实验室(三峡大学),湖北宜昌443002 [2]三峡大学计算机与信息学院,湖北宜昌443002 [3]地质探测与评估教育部重点实验室(中国地质大学(武汉)),湖北武汉430074 [4]中国地质大学(武汉)计算机学院,湖北武汉430074 [5]地理信息系统国家地方联合工程实验室(中国地质大学(武汉)),湖北武汉430074

出　　处：《广西师范大学学报(自然科学版)》2025年第2期107-120,共14页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金(42301492);国家重点研发计划项目(2022YFB3904200,2022YFF0711601);湖北省自然科学基金(2022CFB640);地质探测与评估教育部重点实验室主任基金(GLAB2023ZR01)。

摘　　要：针对平面地质图件中地质构造背景复杂、符号表示多样化而导致识别效果不佳的问题,本文提出一种基于迁移学习和通道先验注意力机制的地质构造识别模型MsAttenEfficientNet。该模型以EfficientNet为主干网络架构,并使用通道先验注意力(channel prior convolution attention,CPCA)模块替换EfficientNet特征提取模块MBConv中的压缩和激励网络(squeeze-and-excitation net,SENet),使模型能够动态地分配通道和空间注意力权重,更准确地捕捉到图像中的重要区域和空间结构;其次对顶层预测模块进行改进,引入Swish激活函数和Dropout层,加强模型的泛化性能;最后使用Adam优化算法提高网络的收敛速度,并利用迁移学习实现特征参数共享。通过在地质构造数据集GeoStr18上进行训练及测试,实验结果表明,MsAttenEfficientNet模型对地质构造的识别精准率为96.92%,召回率为96.89%,F 1分数为96.90%,优于ResNet50、ShuffleNetV2和DenseNet121等主流分类识别模型,可有效用于地质构造识别。To address the issue of poor identification resulting from the complex geological background and diverse symbol representations in plane geological drawings,a geological structure recognition model,named MsAttenEfficientNet,is proposed based on transfer learning and the Channel Prior Convolution Attention(CPCA)mechanism.In this model,EfficientNet is utilized as the backbone network architecture,and the Squeeze and Excitation Net(SENet)in the MBConv feature extraction module of EfficientNet are replaced by the CPCA module.This replacement enables the dynamic allocation of channel and spatial attention weights,allowing the model to more accurately capture important regions and spatial structures in the images.Subsequently,improvements are made to the top prediction module by introducing the Swish activation function and Dropout layers to enhance the model’s generalization performance.Finally,the Adam optimization algorithm is employed to improve the convergence speed of the network,and transfer learning is utilized to achieve feature parameter sharing.Experimental results,obtained from training and testing on the geological structure dataset GeoStr18,demonstrate that the MsAttenEfficientNet model achieves a precision of 96.92%,a recall of 96.89%,and an F 1 score of 96.90%in geological structure recognition,outperforming mainstream classification models such as ResNet50,ShuffleNetV2,and DenseNet121,thus effectively applicable for geological structure recognition.

关 键 词：图像识别 地质构造 EfficientNet网络 通道先验注意力 迁移学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391[自动化与计算机技术—控制科学与工程]