Research on Intelligent recognition of Rock samples based on Deep Learning method

Abstract

Rock sample recognition plays an important role in oil and gas exploration and mineral resources exploration. At present, the main rock sample identification methods are logging, seismic, gravity and magnetism, geochemistry, hand samples and thin slice analysis and so on. In particular, the mainstream thin section analysis requires professionals to analyze the rock slices manually, but manual detection and preparation of rock slices take a long time. In order to reduce the workload and shorten the time period, the image depth learning method is used to establish the model. After the cuttings and core samples are collected by the industrial camera in the mud logging site, the real-time automatic identification and classification of rock samples and the evaluation of oil and gas-bearing properties are realized. In this paper, the file in .JPG format is cut to discard the information that does not contain rock at the edge. Then it was divided into 4* 4, 6*6, 8*8, 10*10 segments. After image enhancement, the transfer learning method is used to train the neural network using Desnet201 in the built-in library of torch, and the Desnet201 network is selected as our neural network framework. Then the last softmax layer is changed into a multi-layer perceptron with three hidden layers, and the improved Desnet201 is used as a model to identify rock types in this paper.