基于ResNet优化的卷积神经网络岩芯图像识别——以广州市典型区域为例  

Convolutional neural network core image recognition based on ResNet optimization——A case study of typical regions in Guangzhou City

作　　者：林禄杰 梁柱 LIN Lu-jie;LIANG Zhu(Guangzhou Urban Planning&Design Survey Research Institute Co.,Ltd.,Guangzhou 510060,Guangdong,China;Collaborative Innovation Center for Natural Resources Planning and Marine Technology of Guangzhou,Guangzhou 510060,Guangdong,China;Guangdong Enterprise Key Laboratory for Urban Sensing,Monitoring and Early Warning,Guangzhou 510060,Guangdong,China)

机构地区：[1]广州市城市规划勘测设计研究院有限公司,广东广州510060 [2]广州市资源规划和海洋科技协同创新中心,广东广州510060 [3]广东省城市感知与监测预警企业重点实验室,广东广州510060

出　　处：《地下水》2024年第6期322-326,共5页Ground water

基　　金：广州市资源规划和海洋科技协同创新中心项目(2023B04J0301,2023B04J0326);浅埋暗挖隧道施工地面坍塌致灾机理与监测预警关键技术研究(RDI2220204037);广东省重点领域研发计划资助(2020B0101130009);广东省城市感知与监测预警企业重点实验室基金项目(2020B121202019)资助。

摘　　要：在人工智能信息化时代,图像识别作为其中的核心技术已深入到各个行业,并对人类生产和生活方式产生了改变。在岩土工程方面也有所应用,勘察外业过程中需要对岩芯信息进行采集,包括现场拍照以及岩性描述。这种传统的勘察记录手段一方面需要大量的人力维持,另一方面信息采集的准确性能以保障。本文通过对人工智能与图像识别算法的研究,提出基于ResNet优化的卷积神经网络模型,逐步实现基于图像的岩性智能识别与勘察外业自动化编录。采集广州市典型岩土体8类,每类1000多张现场照片进行识别试验,所得结果与另一种常用的图像识别算法LeNet进行对比。结果显示,基于ResNet优化的卷积神经网络模型在测试集数据最高准确率达到94.5%,优于LeNet算法的74.5%,从而验证了本模型的先进性,进一步提高岩性识别的效率。In the era of artificial intelligence information, image recognition technology as a core technology has penetrated into various industries and has brought about disruptive changes to human production and lifestyles. The field of geotechnical engineering investigation requires the collection of core photos and the input of corresponding core description information. Through the research of artificial intelligence and image recognition algorithms, this paper selects appropriate models and continuously optimizes them, and gradually realizes image-based lithology Intelligent auxiliary identification and automatic cataloging of survey field. This experiment divides Guangzhou's typical rock and soil layers into 8 categories, each with more than 1,000 sheets. The ResNet model is used for experiments. The results are compared with LeNet, another commonly used image recognition algorithm. The results show that the convolution neural network model optimized based on ResNet has the highest accuracy rate of 94.5% in the test set data, which is better than 74.5% of LeNet algorithm, thus verifying the progressiveness of this model and further improving the efficiency of lithologic identification.

关 键 词：岩性识别 图像识别 ResNet 卷积神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程] P585[天文地球—岩石学]