基于卷积神经网络岩石样本识别方法

Rock Sample Recognition Method Based on Convolution Neural Network

摘要: 岩石种类繁多，人类现已发现的岩石种类有好多种。从现实意义来说，岩石样本识别在油气勘探、水资源勘探、矿物勘探、能源勘察、工程建设中是一项既基础又重要的环节。近年来，随着人工智能技术的发展，采用图像深度学习的方法建立岩石样本自动识别分类模型成为主导，相较于传统岩石分类减少了设备影响与人为因素干扰，更为快速且准确地识别。本研究数据集来自Big Data Mining Race dataset2021，针对该数据集存在岩石种类不均匀、图片大小不一且数据量少的问题，通过图片裁剪、旋转、饱和度调整等图片增强手段，上述问题得以缓解。本研究针对岩石剖面，结合迁移学习，通过冻结与微调预训练模型的部分层，分别建立基于InceptionV3网络、ResNet50网络、VGG16网络的逐步细分串联模型与单个模型，通过准确率、召回率等值判断模型的好坏，最终得出结论。

Abstract: There are many kinds of rocks that have been proven by human beings. From a practical point of view, rock samples identification is a basic and important link in oil and gas exploration, water resources exploration, mineral exploration, energy exploration, and engineering construction. In recent years, with the development of artificial intelligence technology, the method of image deep learning has become the dominant method to establish an automatic identification and classification model of rock samples. Compared with traditional rock classification, it reduces the influence of equipment and human factors, and can identify more quickly and accurately. The data set of this study is from Big Data Mining Race dataset2021. The data set has the problems of uneven rock types, different image sizes and small data volume. The problems can be solved by image enhancement methods such as image cropping, rotation, and saturation adjustment. In this paper, aiming at the rock profile, combined with transfer learning, by freezing and fine-tuning some layers of the pretraining model, a step-by-step subdivision series model and a single model based on the InceptionV3 network, ResNet50 network, and VGG16 network were established respectively. Through the equivalence of accuracy and recall rate, we can judge the quality of the model and finally draw a conclusion.