摘要
岩石图像采集过程难度大、成本高,导致岩性识别的训练样本匮乏。基于岩石图像样本分布导致特征分布不均衡的问题,提出融合注意力的自适应元学习岩石图像分类方法(FA-AML),捕捉岩性分类任务之间的共性,实现少样本场景下的岩性分类;利用融合注意力的残差网络提取最具有区分度的岩石关键特征,缓解特征分布不均衡的问题,并引入元网络自适应更新网络的学习率与正则项系数,提高网络超参数更新自适应性,增强网络适应新类别岩石图像的能力。结果表明,与其他主流元学习方法相比,FA-AML在进行岩石图像分类研究时具有优越的分类性能。
The process of rock image acquisition is difficult and costly,resulting in a lack of training samples for lithology identification.To solve the problem of uneven feature distribution caused by the distribution of rock image samples,an adaptive meta-learning rock image classification method with fusion attention(FA-AML)is proposed to capture the commonalities between lithology classification tasks and realize lithology classification in few-sample scenarios.The residual network with fusion attention is used to extract the most discriminative key features of rock for alleviating the problem of uneven feature distribution,and a meta-network is introduced to adaptively update the learning rate and regular term coefficient of the network for improving the adaptability of network hyperparameter update and enhancing the ability of the network to adapt to new categories of rock images.The experimental results show that the proposed FA-AML method has superior classification performance on rock image classification tasks compared with other mainstream meta-learning methods.
作者
马明刚
潘月梁
彭泽豹
王龙宝
MA Minggang;PAN Yueliang;PENG Zebao;WANG Longbao(Zhejiang Ninghai Pumped Storage Co.,Ltd.,Ninghai 315600,Zhejiang,China;College of Computer and Information,Hohai University,Nanjing 211100,Jiangsu,China)
出处
《水力发电》
CAS
2022年第11期50-54,101,共6页
Water Power
基金
国家自然科学基金资助项目(51779084)。
关键词
岩性识别
图像分类
元学习
自适应
注意力
残差网络
lithology identification
image classification
meta-learning
adaptive
attention
residual network
