基于岩石图像迁移学习的岩性智能识别

Intelligent Lithology Identification Based on Transfer Learning of Rock Images

岩性快速识别具有重要的基础地质研究意义与工程应用价值,本文提出了一种基于岩石图像迁移学习的岩性智能识别方法.首先,利用深度监督目标检测网络(DSOD)对图像中的岩石进行检测,通过获取岩石的位置信息并自动剪裁,建立高质量的岩石图像数据集.然后,结合ResNet网络构建岩石图像深度学习迁移模型,利用残差网络提取岩石特征信息.再利用迁移学习的方法,通过加载预训练权重对模型进行训练,从而实现岩性智能识别.此外,本文采用混淆矩阵、准确率(ACC)、P、R和F1值作为模型准确率的评价指标,对基于ResNet-101和ResNet-50的模型开展了对比分析.结果表明,基于ResNet-101的岩石图像深度学习迁移模型ACC最高可达90.21%,P最高可达91.29%,F1最高可达90.7%.相较于基于ResNet-50的模型,其识别精度更高且对每类岩石识别的稳定性更好.实验和可靠性分析表明,本文提出的岩性智能识别方法具有良好的鲁棒性和泛化性能,可用于地质、测井、交通、水利等工程中的岩性快速智能识别.

An intelligent lithology identification method is proposed based on the deep learning of rock images.Target detection is implemented to extract the position information of rock using the Deeply Supervised Object Detector(DSOD) and to set up a high-quality rock image data set by automatic cutting.Based on the ResNet, a transfer learning model of rock images is constructed to extract rock feature information by using the residual network.Then using transfer learning method, the model is trained for loading pre-training weight, and the intelligent lithology identification can be realized.The model based on the ResNet-50/101 are compared by using the confusion matrix, accuracy(ACC),P,R and F;as the evaluation indexes of the model.Results indicate that the ACC of the transfer learning model based on the ResNet-101 is 90.21%,P is 91.29% and F;is 90.7%.Compared with the ResNet-50 model, the accuracy is higher and the stability of each rock identification is better.The reliability assessment and case study also indicate that the improved lithology identification method is of good robustness and generalization performance, which can be used for quick and intelligent identification of lithology in geology, logging, transportation, water conservancy engineering.