Research on Lithology Recognition Method for Borehole Camera Based on Deep Learning

Lithology is an important parameter in geological survey of hydropower projects, and lithology identification is an important link in geological evaluation. There is a complex nonlinear relationship between lithology and rock composition. With the continuous increase of hydropower engineering exploration data, how to establish appropriate nonlinear relationships from massive lithology data is the focus of lithology identification. Deep neural networks can construct complex nonlinear functions, and convolution has the ability to extract data features. This article uses deep neural networks to identify lithology. We build a deep neural network based on Le-Net5, take the camera pictures inside the borehole as input, and use data enhancement methods such as translation, rotation, and scaling to expand the training samples; adjust the Dropout parameters and add an attention mechanism to reduce model overfitting, through a large number of experiments Contrast to optimize hyper-parameters. The accuracy of this method for classifying the limestone and slate test sets in the study area reached 97.5%, and the accuracy of the validation set reached 96.0%. Experimental results show that adding an attention mechanism and discarding some fully connected layer nodes can prevent model overfitting and improve the accuracy of in-hole camera lithology identification.