Intelligent Identification of Subsurface Rock Mass Fractures Based on YOLOv8 with Efficient Channel Attention (ECA)

Fracture identification is an important aspect of assessing rock mass stability. However, manual evaluation of borehole fracture images can be time consuming and lacks quantitative precision, particularly because of the complex shapes of fractures and the challenges in accurately annotating those images. To address these issues, this study uses the YOLOv8 model as a baseline and enhances its feature extraction capability by adding the efficient channel attention (ECA) module. In addition, the wise-intersection over union (WIoU) loss function is adopted to improve the model ability to precisely capture fracture edges. Experimental results showed that adding either the ECA module or the WIoU loss function to the baseline model effectively improved recognition performance. Among various attention modules, ECA achieved the best results. For the four metrics, namely box mAP@50, box mAP@50–95, mask mAP@50, and mask mAP@50–95, the model with only ECA demonstrated notable improvements, as did the model with only WIoU. Furthermore, the model that combined both the ECA module and WIoU loss achieved increases of 5.1% and 4.0% for box mAP@50 and box mAP@50–95, respectively, and 3.9% and 4.5% for mask mAP@50 and mask mAP@50–95, respectively. These combined improvements were greater than the sum of the individual enhancements, indicating a synergistic effect. Tests on new boreholes demonstrated that the proposed model could effectively identify fractures in borehole TV images. However, owing to differences in data distribution, certain small fractures might still be missed. Further efforts are required in future work to enhance the generalizability of the model.