| 基于岩石图像深度学习的岩性自动识别与分类方法 |
| |
| 引用本文： | 张野,李明超,韩帅.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,2018,34(2):333-342. |
| |
| 作者姓名： | 张野 李明超 韩帅 |
| |
| 作者单位： | 天津大学水利工程仿真与安全国家重点实验室, 天津 300350;国土资源部地质信息技术重点实验室, 北京 100037,天津大学水利工程仿真与安全国家重点实验室, 天津 300350,天津大学水利工程仿真与安全国家重点实验室, 天津 300350 |
| |
| 基金项目： | 本文受国家自然科学基金优秀青年项目（51622904）和面上项目（51379006）及国土资源部地质信息技术重点实验室开放课题（2017321）联合资助. |
| |
| 摘 要： | 岩石岩性的识别与分类对于地质分析极为重要,采用机器学习的方法建立识别模型进行自动分类是一条新的途径。基于Inception-v3深度卷积神经网络模型,建立了岩石图像集分析的深度学习迁移模型,运用迁移学习方法实现了岩石岩性的自动识别与分类。采用此方法对所采集的173张花岗岩图像、152张千枚岩图像和246张角砾岩图像进行了学习和识别分类研究,通过训练学习建立岩石图像深度学习迁移模型,并分别采用训练集和测试集中的岩石图像对模型进行了检验分析。对于训练集中的岩石图像,每组岩石分别用3张图像测试,三种岩石的岩性分类均正确,且分类概率值均达到90%以上,显示了模型良好的鲁棒性;对于测试集中的岩石图像,每组岩石分别采用9张图像进行识别分析,三种岩石的岩性分类均正确,并且千枚岩组图像分类概率均高于90%,但是花岗岩组2张图像和角砾岩组的1张图像分类概率值不足70%,概率值较其他岩石图像低,推测其原因是训练集中相同模式的岩石图像较少,导致模型的泛化能力减小。为了提高识别精确度,对准确率较低的岩石图像进行截取,分别取其中的3张图像加入训练集进行再训练,增加与测试图像具有相同模式的训练样本;在新的模型中,对3张图像进行二次检验,测试概率值均达到85%以上,说明在数据足够的状况下模型具有良好的学习能力。与传统的机器学习方法相比,所提出的岩石图像深度学习方法具有以下优点:第一,模型通过搜索图像像素点提取物体特征,不需要手动提取待分类物体特征;第二,对于图像像素大小,成像距离及光照要求低;第三,采用适当的训练集可获得较好的识别分类效果,并具有良好鲁棒性和泛化能力。
|
| 关 键 词： | 岩石图像 深度学习算法 岩性识别 自动分类 迁移学习 |
| 收稿时间： | 2017/6/1 0:00:00 |
| 修稿时间： | 2017/9/20 0:00:00 |
Automatic identification and classification in lithology based on deep learning in rock images |
| |
| ZHANG Ye,LI MingChao and HAN Shuai.Automatic identification and classification in lithology based on deep learning in rock images[J].Acta Petrologica Sinica,2018,34(2):333-342. |
| |
| Authors: | ZHANG Ye LI MingChao and HAN Shuai |
| |
| Institution: | State Key Laboratory of Hydraulic Engineering Simulation and Safety, Tianjin University, Tianjin 300350, China;Key Laboratory of Geological Information Technology, Ministry of Land and Resources, Beijing 100037, China,State Key Laboratory of Hydraulic Engineering Simulation and Safety, Tianjin University, Tianjin 300350, China and State Key Laboratory of Hydraulic Engineering Simulation and Safety, Tianjin University, Tianjin 300350, China |
| |
| Abstract: | It is important for geology analysis to make identification and classification in lithology. It is a new way to establish the identification model in machine learning. In this research, a transfer learning model of rock images was built based on the Inception-v3 model. It was adapted to process 173 granite images, 152 phyllite images and 246 breccia images to train the transfer learning model. Images in trained data set and in test data set were used to test the model, respectively. 3 images in each group from the trained data set were selected to test the model. There were no identification and classification errors and the all of the probabilities were more than 90%. 9 images in each group from the test data set were also selected to test the model. There were no identification and classification errors. The probabilities of phyllite group were more than 90%. While, the probabilities of 2 images in granite and 1 image in breccia group were less than 70%. It was thought that there were fewer images with similar pattern leading to the bad results. To verify the hypothesis, cut the images with low probabilities and added 3 images to the trained data set in each group to retrain the model. The 3 images with low probabilities were tested in the retrained model and their probabilities were more than 85%. It showed the model had good robustness and generalization if there were enough images. Compared with the traditional machine learning, the proposed method has much strength. First, there is no need to do manual tuning and it processes the data in the model automatically. Second, there is no specific requirement in image pixel, distance and size. At last, the model can have a robust identification and classification result if a suitable trained data set is adopted. |
| |
| Keywords: | Rock images Deep learning Lithology identification Automatic classification Transfer learning |
| 本文献已被 CNKI 等数据库收录！ |
| 点击此处可从《岩石学报》浏览原始摘要信息 |
| 点击此处可从《岩石学报》下载免费的PDF全文 |
|
