Deep learning of rock images for intelligent lithology identification. (September 2021)

• Record Type: Journal Article
• Authors: Xu, Zhenhao; Ma, Wen; Lin, Peng; Shi, Heng; Pan, Dongdong; Liu, Tonghui
• Is Part Of: Computers & geosciences. Volume 154(2021)
• Publication Date: 2021-09
• Subjects: Deep learning -- Rock images -- Intelligent detection -- Lithology identification; Environmental policy -- Periodicals; 550.5
• DOI: 10.1016/j.cageo.2021.104799
• Languages: English
• ISSNs: 0098-3004

Abstract

Abstract: An intelligent lithology identification method is proposed based on the deep learning of rock images. The lithology information and position information in rock images can be predicted using the Faster R–CNN architecture through the RPN proposal generation algorithm and the Fast R–CNN detector. To obtain more rock features, the rock detection model is built on the ResNet structure, and the residual learning is used to retain as much as possible detailed information in the original input image. The four-step alternating training is used to fine-tuned end-to-end, and the prediction results are optimized by the cross-entropy loss and the regression loss. To speed up the model and improve the identification accuracy, data augmentation and pre-training are used to train the model. The mAP, P, R and F 1 score are used as evaluation indexes of the accuracy, and the Faster R–CNN model is compared with the YOLO v4 model. Results indicate that the mAP of the rock detection model based on the Faster R–CNN is 99.19% and the F 1 score is 96.6%. Compared with the YOLO v4 model, the accuracy is higher and the identification ability is more stable. The proposed rock detection model has good identification ability for different rocks in rock images, and the model is of good robustness and generalization performance, which is suitable for rapid intelligent lithology identification in practical geological and logging engineering. Highlights: A rock detection model is built on the Faster R–CNN for position and lithology prediction. The ResNet-50 is applied to build the Fast R–CNN detector and to extract more rock deep features. Case study is carried out to verify the intelligent lithology identification method.