Detection of Rock Discontinuities in Borehole Images Based on a Deep Learning Method

Rock discontinuity has an important influence on the stability and permeability of the rock mass,thereby significantly determining the stability of deep underground engineering.This paper proposes a method based on a deep learning algorithm (You Only Look Once version 4,YOLO v4) for the identification and positioning of rock discontinuities and the calculation of geometric parameters using borehole images obtained by downhole television technology.First,image data is collected and preprocessed.For the tunnel project,the Intelligent Drilling Optical Imager equipment is used to collect images of borehol No.4 and No.6,and the images were processed to seperate those containing rock discontinuities with annotations to create a Ground truth dataset,with 70% of the dataset randomly selected under data enhancement as training data,10% as verification data and 20% as test data.The CSPDarkNet53 was then used as the feature extraction network to build the YOLO v4 model,with optimal parameters determined through trial and error for model training.The test set was designed to generate the P-R (Precision-Recall) curve to test the final model training effect,and the result shows that the average precision of the P-R curve reaches 0.87,indicating that the training result of YOLO v4 is better.Finally,the Canny algorithm was used to obtain the edges of the rock discontinuities by fitting the upper,middle and lower sine functions to the localised structural surface,allowing the four geometric parameters (dip direction,dip angle,depth and aperture) to be calculated.