摘要: 随着计算机硬件性能的大幅提升和人工智能研究的快速发展，人工智能识别技术已应用于医学、教育、工业、商业、勘探等众多行业专业领域。在地质勘探中，岩石样品的分类识别是地质分析的重要环节。然而，传统的人工分类方法面临着较高的经济和时间成本，且容易受到主观判断的影响而影响识别结果。为了避免这些问题，研究了基于计算机的人工智能识别技术在岩石图像识别中的具体应用以及不同参数造成的性能差异。实验中使用的样本数据为测井现场工业相机采集的岩屑和岩心样品，共计350幅图像，涵盖7种类型的岩石样品。本文将这350幅岩石图像按一定比例划分为训练集、验证集和测试集。由于数据集较小且各类岩石的数量各不相同，因此在训练之前采用数据增强技术对原始数据进行扩充，扩充过程中均衡了不同类型岩石样品的数量。在训练阶段，利用开源深度学习框架Pytorch构建多层卷积神经网络，对训练集进行学习和分类训练，测试中最佳准确率达到82.86%。然而，对于拉伸图像的识别率较低，需要进一步优化网络结构。基于实验结果总结了神经网络训练的优化方向，包括提升数据集质量、优化网络结构、调整训练策略等。为了尽可能直观地展示实验数据，实验中使用PlotNeuralNet、Python matplotlib等数据可视化工具进行具体的可视化工作，并从数据与图像相结合的角度对实验数据及模型效果进行分析。

Abstract: With substantial improvement of computer hardware performance and rapid development in artificial intelligence research, artificial intelligence recognition technology has been applied in many industries, such as medicine, education, industry, commerce, exploration and other professional fields. In geological exploration, the classification and identification of rock samples is an important link in geological analysis. However, the traditional manual classification faces high economic and time costs, and is vulnerable to subjective judgment which may affect the recognition result. In order to avoid these problems, research is carried out on the specific application of computer-based artificial intelligent recognition technology in recognition of rock images and the performance difference caused by different parameters. The sample data used in the experiment are lithic fragments and core samples capture by industrial cameras at the well logging site, with a total of 350 images which cover 7 types of rock samples. In this paper, the 350 rock images are divided into training set, validation set and test set based on certain proportion. Since the data set is too small and the quantity of each type of rocks varies, data augmentation technology is used to expand the original data before training. During the expansion, the quantity of different types of rock samples are balanced. At the training stage, the open-source deep learning framework Pytorch is used to construct a multi-layer convolutional neural network for learning and classification training on the training set. The best accuracy rate in the tests reaches 82.86%. However, the recognition rate is poor for stretched images, which required further optimization of the network structure. Based on the experimental results, the optimization orientations of neural network training are summarized, including improvement of data set quality, optimization of network structure, and adjustment of training strategies. In order to display the experimental data as intuitively as possible, data visualization tools such as PlotNeuralNet and Python matplotlib are used in the experiment for specific visualization work, after which the experimental data and model effect are analyzed from the perspective with combination of data and images.