Abstract: [Objective]Rock joint detection in tunnel engineering is a critical component for maintaining structural stability. Current tunnel inspection methods, influenced by human subjectivity, suffer from high rates of missed and false detections, as well as limited capabilities in global localization and capture of subtle joints. [Methods]To address these issues, this paper proposes the GeoLA-YOLO algorithm—a high-efficiency rock joint recognition system for tunnel engineering. By incorporating a Convolutional Block Attention Module (CBAM) into the backbone network, the algorithm enhances its ability to capture subtle feature information, effectively resolving the challenge of extracting fine details. Furthermore, through improvements to the head architecture, the model achieves enhanced precision in locating and identifying subtle joints, thereby addressing the issue of inaccurate global positioning. [Results]Experimental results on our self-built VOC (Visual Object Classes) dataset demonstrate that the optimized algorithm maintains lightweight performance while improving mAP@0.5, mAP@0.5-0.95, Recall, and F1 metrics by 4.3%,9.6%,5.0%, and 5.5% respectively compared to the original algorithm, validating the model's effectiveness. In public datasets, the improved model shows 6.2% and 5.2% higher mAP@0.5, mAP@0.5-0.95 performance than the baseline algorithm, confirming GeoLA-YOLO's robustness.