Abstract
Rock core logging plays a crucial role in acquiring geological and geotechnical information. However, conventional manual core logging is labor-intensive and requires extensive expertise. To address these challenges, this study proposes a deep learning-based object detection model called Att-YOLO, to automate rock core classification and localization tasks. The proposed model incorporates the convolutional block attention module (CBAM) into the backbone and integrates the YOLOv5 PAN-FPN structure in the neck. A database of corebox images is compiled from Hong Kong ground investigation reports. Furthermore, transfer learning is employed during the training process to expedite model convergence. The performance of Att-YOLO model is evaluated on a test set, achieving the highest mean average precision (mAP) compared to other prevalent algorithms. Further visualization experiments are conducted to investigate the causes of errors and feature processing capabilities across varying model depths. Ablation studies are also implemented to demonstrate the effectiveness of the CBAM module. Finally, the extra corebox images from various locations in Hong Kong are collected and used to verify the robustness of the Att-YOLO model. The proposed model exhibits strong generalization ability for multi-scale objects from different projects, proving its effectiveness in processing rock core data.