为增强地铁保护区内钻机施工识别的及时性和准确性，对比分析了Faster R-CNN(Faster-Regions with CNN Features), SSD(Single Shot MultiBox Detector) 和YOLO(You Only Look Once)3种图像识别算法模型的优缺点和适用场景，构建了结合马赛克数据增强和学习率余弦退火算法的地铁保护区钻机检测方法，分析了权重衰减系数等参数和算法对于识别准确率和帧率等的敏感性。结果表明：YOLO模型系列中的YOLOv4模型对于钻机识别的平均准确率达到了94.03%，帧率为8.9fps，精确率、召回率及调和平均数也均超过了Faster R-CNN, SSD和YOLOv3模型，并且在同时使用马赛克数据增强和学习率余弦退火算法时平均准确率达到最高。由此说明，YOLOv4模型在钻机识别中适用性较好，可以有效实现对监控影像中钻机图像的自动识别、检测和预警，为实时监测、快速处置保护区违规施工提供技术支撑。

In order to enhance the timeliness and accuracy of drilling construction identification in subway protection zone, the advantages, disadvantages and applicable scenarios of three image recognition algorithm models, such as Faster R-CNN(Faster-Regions with CNN Features), SSD(Single Shot MultiBox Detector) and YOLO(You Only Look Once), were compared and analyzed. A drilling detection method in subway zone combining Mosaic data enhancement and cosine annealing algorithm of learning rate was constructed. The sensitivity of parameters and algorithms such as weight attenuation coefficient to recognition accuracy and frame rate were analyzed. Experimental results showed that the average accuracy rate of YOLOv4 model in the Yolo model series for drilling recognition was 94.03%, and the frame rate was 8.9fps; the accuracy rate, recall rate and harmonic mean were also higher than those of Faster R-CNN, SSD and YOLOv3 models, and the average accuracy rate reached the highest value when Mosaic data enhancement and learning rate cosine annealing algorithm were used simultaneously. All these suggest that YOLOv4 model has a high applicability in drilling identification, which can effectively automatically identify, detect and warn the drilling machine images in the monitoring images, and provide technical support for real-time monitoring and rapid disposal of illegal construction in subway protection zone.