基于深度学习与数据增强技术的小样本岩石分类

在油气勘探中,利用深度学习技术对岩石进行识别与分类能极大提高工作效率。岩石采样并制作样本图像费时费力,因此岩石样本通常较少。有鉴于此,基于深度学习技术设计一个新的神经网络模型MyNet,该模型能对小样本进行学习并完成岩石样本的分类。使用数据增强技术通过Python编程将314张岩石样本扩充成28272张图像,为了充分利用现有数据,取其中的27384张作为训练集,剩余888张作为测试集。将数据分别导入MyNet、ResNet50、Vgg16进行训练和测试。实验结果表明,加载、不加载预训练参数的ResNet50、Vgg16的岩石分类结果因受有无迁移学习影响会有所不同；MyNet的总体分类准确率为75.6%,均优于有无迁移学习的ResNet50、Vgg16,且其结构复杂度与训练代价明显低于比较模型,说明新模型应用于小样本的岩石分类可行有效且经济安全,更容易推广应用。

Small rock samples classification based on deep learning and data enhancement technologies

In oil and gas exploration, using deep learning technology to identify and classify rocks can greatly improve work efficiency. Rock sampling and making sample images are time-consuming and laborious, hence there are usually few rock samples. In view of this, a new neural network model MyNet is designed based on deep learning technology. The model can learn small samples and complete the classification of rock samples. Using data enhancement technology, 314 rock samples are expanded into 28272 images by Python programming. In order to make full use of the existing data, 27384 of them are taken as the training set and the remaining 888 as the test set, and then import the data into MyNet, ResNet50 and Vgg16 for training and testing. The experimental results show that the rock classification results of ResNet50 and Vgg16 with and without pre training parameters are different due to the presence or absence of migration learning. The overall classification accuracy of MyNet is 75.6%, which is better than ResNet50 and Vgg16 with or without transfer learning, and its structural complexity and training cost are significantly lower than the comparison model. Therefore, the new model is feasible, effective, economical and safe for small rock samples classification, and is easier to be popularized and applied.