| 王孝臣,夏庆霖,徐杨青,杨旭,周润杰,冷帅,周长春,刘荻,马宏超.基于轻量级神经网络的岩石图像智能识别方法研究[J].矿产勘查,2026,17(3):556-565 |
| 基于轻量级神经网络的岩石图像智能识别方法研究 |
| Research on intelligent recognition method of rock images based on lightweight neural networks |
| 投稿时间：2025-04-13 |
| DOI：10.20008/j.kckc.202603012 |
| 中文关键词: 岩性识别 岩石图像 轻量级神经网络 注意力机制 迁移学习 |
| 英文关键词: lithology recognition rock images lightweight neural networks attention mechanism transfer learning |
| 基金项目:本文受山西工程技术学院博士科研启动基金(2024QD-16)、湖北省自然科学基金联合基金(2023AFD225)和新疆维吾尔自治区重大科技专项课题(2024A03001-1)联合资助。 |
| 作者 | 单位 | | 王孝臣 | 山西工程技术学院地球科学与工程系,山西阳泉 045000 | | 夏庆霖 | 新疆维吾尔自治区地质研究院,新疆乌鲁木齐 830091 | | 徐杨青 | 中煤科工集团武汉设计研究院有限公司,湖北武汉 430064
煤炭科学研究总院,北京 100013 | | 杨旭 | 东北大学,辽宁沈阳 110167 | | 周润杰 | 湖北省地质调查院,湖北武汉 430034
资源与生态环境地质湖北省重点实验室,湖北武汉 430034 | | 冷帅 | 湖北省地质调查院,湖北武汉 430034
资源与生态环境地质湖北省重点实验室,湖北武汉 430034 | | 周长春 | 山西工程技术学院地球科学与工程系,山西阳泉 045000 | | 刘荻 | 新疆维吾尔自治区地质研究院,新疆乌鲁木齐 830091 | | 马宏超 | 新疆维吾尔自治区地质研究院,新疆乌鲁木齐 830091 |
|
| 摘要点击次数: 66 |
| 全文下载次数: 22 |
| 中文摘要: |
| 岩性智能识别技术及其应用在地质调查、矿产勘查、工程地质等众多领域中发挥着越来越重要的作用。为解决野外岩性识别主要依赖地质人员经验且智能化程度较低等问题，本文采用一种 PP-LCNet轻量级深度卷积神经网络模型对岩石岩性进行智能识别。首先，搜集建立“二长花岗岩”“花岗闪长岩”“白云石大理岩”“方解石大理岩”“长石石英岩”“煤”“石灰岩”“石英砂岩”“长石砂岩”等不同种类岩石图像数据集，并对图像进行分割与预处理；选择预先训练过的 PP-LCNet轻量级网络模型，并结合 SE(Squeeze-and-Excitation)通道注意力机制模块，对 1890张岩石训练集图像进行迁移训练，并利用验证集不断调整模型参数，得到轻量且高精度的图像识别模型；利用测试数据集进行模型测试，评估模型性能，显著提高了模型的泛化能力，很好解决了岩性识别模型算法精度与速度难以平衡的难点。相较于 Mobile-NetV2、MobileNetV3、ResNet50、Xception65以及 ShuffleNetV2等网络模型，PP-LCNet模型整体评价指标都高于 80%。该模型体积小，仅占用 12.9 MB的空间，单张岩石图片识别时间仅约为 2.85 s，显示出更高的准确率和快速的识别速度。因此，该模型在岩性识别方面具有较大优势和应用前景，能够有效提高岩性智能识别的精度和效率。 |
| 英文摘要: |
| Lithology intelligent identification technology and its application play an increasingly important rolein many fields such as geological survey, mineral exploration and engineering geology. In order to solve the problemthat field lithology identification mainly depends on the working experience of geological personnel and the intelli-gence degree is low, a lightweight deep convolutional neural network model called PP-LCNet was adopted for intel-ligent recognition of rock lithology. Firstly, collect and establish six different types of rock image datasets, namelymonzonitic granite, granodiorite, dolomite marble, calcite marble, arkosite, coal, limestone, quartz sandstone and ar-kose, and segment and preprocess the images; then, a pre-trained PP-LCNet lightweight network model was se-lected, and combined with the SE (Squeeze-and-Excitation) channel attention mechanism module to transfer andtrain 1890 rock training set images. The model parameters were continuously adjusted using the validation set to ob-tain a lightweight and high-precision image recognition model; Finally, the model was tested using a test dataset toevaluate its performance, significantly improving its generalization ability. This effectively solved the difficulty ofbalancing the accuracy and speed of the lithology recognition model algorithm. Compared to other network modelssuch as MobileNetV2, MobileNetV3, ResNet50, Xception65, and ShuffleNetV2, the overall evaluation index of thePP-LCNet network model was higher than 80%. The model has a small size, occupying only 12.9 MB of space, andthe recognition time for a single rock image is only about 2.85 s, demonstrating higher accuracy and fast recognitionspeed. Therefore, this model has significant advantages and application prospects in lithology recognition, and caneffectively improve the accuracy and efficiency of intelligent lithology recognition. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|