-
摘要:
岩石薄片图像中包含了大量肉眼无法观察到的地质特征信息，对岩石薄片图像的岩性识别结果为后续的石油勘探和开发奠定了基础。针对岩性识别数据集不均衡、识别模型参数多等问题，提出一种改进的轻量化MobileViT模型，该模型针对涵盖了90%以上常见岩性的岩石薄片图像进行建模分析。首先，为使模型更好地学习到每类岩石薄片图像中所包含的独特特征，对数据集进行数字增加。其次，使用GELU替换MobileViT中MV2模块中常规ReLU6，从而作为该模块的激活函数，有效解决神经元死亡的问题，提升模型的收敛速度。最后，划分训练集和测试集，使用余弦退火算法自动更新学习率，以迁移学习加速训练过程，实现岩石薄片图像中针对岩性的自动识别。实验结果表明，改进后的MobileViT对岩性识别的准确率达82.9%，模型的参数仅为7.66M，通过实例验证该算法具有较好的鲁棒性。
Abstract:The rock thin−section images contain a large amount of geological feature information that cannot be observed with the naked eye. The lithology identification of rock thin−section images lays the foundation for subsequent oil exploration and production. Aiming at the problems of unbalanced lithology identification data set and many identification model parameters, an improved lightweight MobileViT model is proposed to model and analyze the rock slice images covering more than 90% of common lithology. First, to enable the model to better learn the unique features contained in each type of rock slice image, adding numbers of the dataset set is performed on the image. Secondly, use GELU to replace the ReLU6 of the MV2 module in MobileViT as the activation function of the module, which effectively solves the problem of neuron death and improves the convergence speed of the model. Finally, the training set and the test set are divided, the cosine annealing algorithm is used to automatically update the learning rate, and the transfer learning is used to speed up the training process, so as to realize the automatic identification of rock slice images. The experimental results show that the accuracy of the improved MobileViT for lithology identification is 82.8%, and the model parameters are only 7.66M, which has good robustness.
-
Key words:
- rock thin section /
- MobileViT /
- lithology identification /
- cosine annealing /
- light weight
-
-
表 1 岩石薄片图像种类及数量分布
Table 1. Types and quantity distribution of rock slice images
岩石
类型岩性 种类 数量 沉积岩 火山碎屑岩、砂岩、泥页岩和粉砂岩、灰岩、
白云岩、硅质岩、蒸发岩、其他内源沉积岩
28
699火成岩 超基性（超镁铁质）岩、基性岩、中性岩、
酸性岩、碱性岩及相关岩石
40
963变质岩 糜棱岩、角岩、矽卡岩、大理岩、蛇纹岩、
云英岩、板岩、千枚岩、片岩、片麻岩、变粒岩、
斜长角闪岩、麻粒岩、榴辉岩、混合岩、
碎裂岩、含变余构造的变质岩
40
972
下载: 导出CSV
表 2 训练集的准确率与训练时长对比
Table 2. Comparison of training set accuracy and training time
模型 最高准确率 是否迁移学习 训练时长/min ResNet50 90.2% 否 905.2 ResNeXt50 88.1% 否 988.9 VGG 62.3% 否 1034.0 MobileViT 100% 是 505.4 Improved_MobileViT 100% 是 422.3
下载: 导出CSV
表 3 测试集的准确率与模型参数对比
Table 3. The accuracy of the test set compared with the model parameters
模型 准确率 参数量 ResNet50 59.8% 97.49M ResNeXt50 56.9% 95.48M VGG 40.1% 527.79M MobileViT 80.2% 7.66M Improved_MobileViT 82.9% 7.66M
下载: 导出CSV
-
[1] Dosovitskiy A, Beyer L, Kolesnikov A, et al. 2020. An image is worth 16×16 Words: Transformers for image recognition at scale[C]//IEEE Conference on Computer Vision and Pattern Recognition (CVPR). doi:10.48550/arXiv.2010.11929.
[2] Hendrycks D, Gimpel K. 2016. Bridging nonlinearities and stochastic regularizers with gaussian error linear Units[J]. abs, 1606.08415.
[3] He K, Zhang X, Ren S, et al. 2016. Deep Residual Learning for Image Recognition[C]//IEEE Conference on Computer Vision and Pattern Recognition (CVPR). doi:10.48550/arXiv.1512.03385.
[4] Hinton G. 2010. A practical guide to training restricted boltzmann machines[J]. Momentum, 9(1): 926−947.
[5] Loshchilov I, Hutter F. 2016. SGDR: Stochastic Gradient Descent with Warm Restarts[J]. CoRR, 2016, arXiv preprint arXiv: 1608.03983.
[6] Marmo R, Amodio S, Tagliaferri R, et al. 2005. Textural identification of carbonate rocks by image processing and neural network: Methodology proposal and examples[J]. Computers & geosciences, 31 (5): 649-659.
[7] Mehta S, Rastegari M. 2021. MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer[C]//IEEE Conference on Computer Vision and Pattern Recognition (CVPR). doi:10.48550/arXiv.2110.02178 .
[8] Pan S J, Qiang Y. 2010. A Survey on Transfer Learning[J]. IEEE Transactions on Knowledge and Data Engineering, 22(10): 1345-1359.
[9] Sandler M, Howard A, Zhu M, et al. 2018. Mobile NetV2: Inverted residuals and linear bottlenecks[C]//IEEE. Conference on Computer Vision and Pattern Recognition (CVPR). doi:10.48550/arXiv.1801.04381.
[10] Simonyan K, Zisserman A. 2014. Very deep convolutional networks for large-Scale image recognition[J]. Computer Science, arXiv preprint arXiv: 1409.1556.
[11] Singh N, Singh T N, Tiwary A, etal. 2010. Textural identification of basaltic rock mass using image processingand neural Network[J]. Computational Geosciences, 14: 301−310. doi: 10.1007/s10596-009-9154-x
[12] 陈国军, 姜朕, 尹冲, 等. 2023. 基于深度学习的岩石孔隙分割方法[J]. 计算机与数字工程, 51(5): 1157−1162. doi: 10.3969/j.issn.1672-9722.2023.05.032
[13] 盖荣丽, 蔡建荣, 王诗宇, 等. 2021. 卷积神经网络在图像识别中的应用研究综述[J]. 小型微型计算机系统, 42(9): 1980−1984. doi: 10.3969/j.issn.1000-1220.2021.09.030
[14] 付光明, 严加永, 张昆, 等. 2017. 岩性识别技术现状与进展[J]. 地球物理学进展, 32(1): 26−40. doi: 10.6038/pg20170104
[15] 宫清顺, 黄革萍, 孟祥超, 等. 2012. 三塘湖盆地火山岩岩性识别方法[J]. 中国石油勘探, 17(3): 37−41, 6. doi: 10.3969/j.issn.1672-7703.2012.03.006
[16] 李嘉璇. 2017. TensorFlow技术解析与实战[M]. 人民邮电出版社: 314.
[17] 刘强, 滕奇志, 何海波. 2022. 岩石薄片多聚焦图像融合的深度学习算法研究[J]. 电子测量技术, 45(3): 125−130.
[18] 赖文, 蒋璟鑫, 邱检生, 等. 2020. 南京大学岩石教学薄片显微图像数据集[J]. 中国科学数据(中英文网络版), 5(3): 26−38.
[19] 马泽栋, 马雷, 李科, 等. 2022. 基于岩石图像深度学习的多尺度岩性识别[J]. 地质科技通报, 41(6): 316−322.
[20] 熊越晗, 刘东燕, 刘东升, 等. 2021. 基于岩样细观图像深度学习的岩性自动分类方法[J]. 吉林大学学报(地球科学版), 51(5): 1597−1604.
[21] 徐述腾, 周永章. 2018. 基于深度学习的镜下矿石矿物的智能识别实验研究[J]. 岩石学报, 34(11): 3244−3252.
[22] 许振浩, 马文, 林鹏, 等. 2021. 基于岩石图像迁移学习的岩性智能识别[J]. 应用基础与工程科学学报, 29(5): 1075−1092.
[23] 张珂, 冯晓晗, 郭玉荣, 等. 2021. 图像分类的深度卷积神经网络模型综述[J]. 中国图象图形学报, 26(10): 2305−2325. doi: 10.11834/jig.200302
[24] 张野, 李明超, 韩帅. 2018. 基于岩石图像深度学习的岩性自动识别与分类方法[J]. 岩石学报, 34(2): 333−342.
[25] 周永章, 陈烁, 张旗, 等. 2018. 大数据与数学地球科学研究进展——大数据与数学地球科学专题代序[J]. 岩石学报, 34(2): 255−263.
[26] 周永章, 左仁广, 刘刚, 等. 2021. 数学地球科学跨越发展的十年: 大数据、人工智能算法正在改变地质学[J]. 矿物岩石地球化学通报, 40(3): 556−573, 777.
[27] 周志华. 2016. 机器学习[M]. 北京: 清华大学出版社.
-
图(10)
表(3)
计量
- 文章访问数: 709
- PDF下载数: 131
- 施引文献: 0