基于 YOLOv5 和 UNet++ 深度学习方法的岩体裂隙智能定位与识别

针对传统岩体裂隙地面调查效率低、高陡边坡环境裂隙调查困难、裂隙图像识别精度低等问题，文章利用地理信息技术中无人机航拍采集的图像，结合YOLOv5和UNet++深度学习网络的方法进行岩体裂隙智能定位与识别。首先利用YOLOv5进行岩体裂隙结构面的检测与定位，然后基于UNet++深度学习网络对检测到的裂隙结构面进行图像识别，并对野外环境下裂隙图像识别的鲁棒性和准确率进行评价。结果表明，在相同样本条件下，YOLOv5的检测精确率达到85.41%，高于Faster R-CNN（faster region-based convolutional neural network）、YOLOv3，UNet++的识别精确率达到82.33%，高于FCN（fully convolutional networks）、UNet。基于YOLOv5和UNet++的深度学习方法能够相对高效准确地实现岩体裂隙的检测定位与识别，可满足野外环境下裂隙批量化快速识别的要求。

Aiming at the problems of low efficiency in traditional ground survey of rock mass fractures, difficulty in environmental fracture investigation of high-steep slope, and low accuracy of fracture image recognition, this paper employs images collected by unmanned aerial vehicle (UAV) aerial photography in geographic information technology as well as the YOLOv5 and UNet++ deep learning methods to conduct intelligent positioning and identification of rock mass fractures. YOLOv5 is used to detect and locate the fracture plane of rock mass, and then image recognition of the detected fracture plane is carried out based on the UNet++ deep learning network to further evaluate the robustness and accuracy of fracture image recognition in the field environment. The results show that under the same sample conditions, the detection accuracy of YOLOv5 reaches 85.41%, which is higher than that of faster region-based convolutional neural network (Faster R-CNN) and YOLOv3, and the recognition accuracy of UNet++ reaches 82.33%, which is higher than that of fully convolutional networks (FCN) and UNet. The deep learning algorithm based on YOLOv5 and UNet++ can realize the detection, positioning and identification of rock mass fractures efficiently and accurately, which can meet the requirements of rapid batch identification of fractures in the field environment.