Rock Pore Network Modeling Based On Deep Learning

The development of deep learning provides new ideas for the research of target detection.The powerful feature extraction capability brought by deep neural networks abstracts the low-level features in the original input data into high-level features,which improves the target detection capabilities.As an important part of rock component research,rock pores have a significant impact on oil and gas exploration and development.The traditional rock pore identification is mainly artificial,and there are problems such as long recognition period and high error rate.Sandstone pore images of tight reservoirs are experimental data,and three types of pores are detected using deep learning.In this paper,the Faster R-CNN network is applied to the detection of rock pores.On the basis of the implementation of the algorithm model,first,a basic experiment based on VGG16 is carried out.Through the analysis of the basic experimental results,small target misdetections and scale differences in rock pore images are found problem.In order to further improve the accuracy of Faster R-CNN for rock pore detection,this paper uses three optimization schemes.One is to change the feature extraction layer in the CNN network to ResNet-101,which improves the stability of the network;the second is to design the RPN network.The structure improves the detection accuracy of the network for small targets;the third is to solve the problem of inconsistent proportion of positive and negative samples by adding OHEM algorithm.Through the improvement of the above three aspects of the network,the overall detection accuracy of the network for small and multi-targets in rock pore images has been improved.