地学前缘 ›› 2024, Vol. 31 ›› Issue (4): 87-94.DOI: 10.13745/j.esf.sf.2024.5.6

• 深度学习与图像识别 • 上一篇    下一篇

基于数据增强和集成学习的矿物图像识别

王琳1(), 季晓慧1,*(), 杨眉2, 何明跃2, 张招崇3, 曾姗1, 王玉柱1   

  1. 1.中国地质大学(北京) 信息工程学院, 北京 100083
    2.中国地质大学(北京) 国家岩矿化石标本资源库, 北京 100083
    3.中国地质大学(北京) 地球科学与资源学院, 北京 100083
  • 收稿日期:2023-08-30 修回日期:2024-02-27 出版日期:2024-07-25 发布日期:2024-07-10
  • 通信作者: * 季晓慧(1977—),女,博士,副教授,主要从事人工智能应用研究。E-mail: xhji@cugb.edu.cn
  • 作者简介:王 琳(1999—),女,硕士研究生,主要从事深度学习、矿物图像识别研究。E-mail: wanglin1911@email.cugb.edu.cn
  • 基金资助:
    国家科技资源共享服务平台——国家岩矿化石标本资源库子项目(NCSTI-RMF20230107)

Mineral identification based on data augmentation and ensemble learning

WANG Lin1(), JI Xiaohui1,*(), YANG Mei2, HE Mingyue2, ZHANG Zhaochong3, ZENG Shan1, WANG Yuzhu1   

  1. 1. School of Information Engineering, China University of Geosciences (Beijing), Beijing 100083, China
    2. National Mineral Rock and Fossil Specimens Resource Center from MOST, China University of Geosciences (Beijing), Beijing 100083, China
    3. School of Earth Sciences and Resources, China University of Geosciences (Beijing), Beijing 100083, China
  • Received:2023-08-30 Revised:2024-02-27 Online:2024-07-25 Published:2024-07-10

摘要:

矿物识别是地质学研究的一个重要部分,对于资源勘探、岩石分类和地质环境监测都有着重要的意义。然而,传统方法通常依赖人的经验进行主观判断,并且效率低下。近年来,已有许多研究将深度学习的图像分类技术应用于矿物识别,以客观快速地识别矿物,这些研究都取得了一定的成果,但可识别矿物种类有限且精度需要进一步提升。为此本文首先解决了矿物数据集图像数据样本分布不平衡问题,对数据集中矿物图像较少的11个矿物类别采用DCGAN生成矿物图像进行数据增强,对比选择效果更好的方案对数据集进行扩充。其次,为了得到更可靠、精确度更高的识别模型,将ImageNet上表现较好的ResNet、RegNet、EfficientNet和Vision Transformer模型迁移到本文使用的矿物数据集上。针对训练好的基模型排列组合得到11个子模型,分别使用平均软投票法和加权软投票法两种方法进行集成,得到22个集成模型并对其训练得到识别结果,对比22个集成模型的结果选择出精度最高的集成模型。实验结果表明:使用DCGAN进行数据增强,在不同的模型上平均提升了3.12%的准确率,充分证明了DCGAN数据增强的有效性;在所有集成模型中,使用加权软投票法的模型表现较好,其中精度最高的是利用4个基分类模型进行加权软投票得到的集成模型,在扩充后的36种常见矿物数据集上达到了87.47%的准确率。

关键词: 矿物识别, 深度卷积生成对抗网络, 数据增强, 集成学习

Abstract:

Mineral identification as a crucial aspect of geosciences is of great importance to resource exploration, rock classification, and geological monitoring. However, traditional methods are inefficient as they often rely on human experience and subjective judgment. In recent years deep learning-based image classification has been used for accurate and rapid mineral identification. While these studies have achieved certain results, the number of identifiable mineral types are limited and the identification accuracy need to be further improved. This paper aims to address the issue of uneven distribution of mineral image samples in a mineral dataset on 36 common minerals. DCGAN is first used to generate images for data augmentation focusing on the 11 minerals with low sample counts, and the best set of images is selected, by comparison, to expand the dataset. Next, to obtain a more reliable and precise identification model, ResNet, RegNet, EfficientNet, and Vision Transformer models with better performance on ImageNet are transferred to the mineral dataset. Based on the permutations of the trained base models, 11 ensemble models are obtained, with which 24 identification results are obtained using two voting methods, average and weighted soft voting. These results are then compared to select the one with the highest accuracy. The experimental results demonstrated that data augmentation using DCGAN improved the model accuracy by 3.12% averaged over all models. Among the ensemble models, weighted soft voting performed better and achieved the highest accuracy of 87.47% on the augmented dataset.

Key words: mineral identification, deep convolutional generative adversarial networks, data augmentation, ensemble learning

中图分类号: