Automatic Estimation Of Rock Quality Designation Based On An Improved YOLOv5

Rock quality designation (RQD) characteristics for assessing the degree of rock mass fracture make it a key parameter in rock grading or other rating systems. Traditional core characterization relies on subjective manual visual inspection by geologists. Currently, convolutional neural networks are used in borehole images to classify intact and nonintact cores in core rows for automatic RQD estimation. Classification networks cannot predict the exact locations of the intact cores, and drill core characterization is not intuitive. Alternatively, an attention mechanism combining channel and spatial attention modules is proposed to improve the YOLOv5 algorithm for drill core characterization. The model was trained on 657 artificial core tray images generated by the developed preprocessor to accurately predict the bounding boxes of the intact cores on the row centerline, and the automatic RQD calculation of the row was implemented with the developed postprocessing program. Our method performed RQD estimation on 602 new granite rows and 180 new quartz sandstone rows, with average error rates of 1.27% and 1.12%, respectively. It processed 50 m of cores on average in 1 s on a GPU. Furthermore, this method provides an innovative method for automatically processing and quantifying geological image databases.