露天矿边坡裂隙智能识别与信息解算

Intelligent identification and information calculation of slope crack in open-pit mine

节理裂隙是影响露天矿边坡稳定性的重要因素之一,随着图像处理技术以及机器视觉技术的发展,采用智能算法进行识别已成为热点.为快速获取节理裂隙几何信息,通过ResNet系列算法对U-net的骨架构网络进行改进,提出了一种露天矿边坡裂隙识别及几何参数解译方法.利用无人机综合考虑视角、距离、重叠率以及飞行速度等因素对露天矿边坡裂隙航拍获取高清图像,使用全局阈值分割技术进行预处理,并运用随机旋转、随机亮度及对比度调整等方式进行数据增广形成裂隙图像数据集;采用残差网络(ResNet)对U-Net网络的骨架构网络进行改进,提出基于改进U-net网络的边坡裂隙识别模型,基于像素二分类问题采用准确率(Accuracy)、交并比(IoU)和F1分数(F1 Score)作为评价指标,结合裂隙图像数据集对提出模型进行训练和评估,输出裂隙二值图,并与传统裂隙识别方法识别结果进行对比;对裂隙二值图进行裂隙几何参数信息解算,获得裂隙长度、宽度统计分布规律和参数.结果表明:ResNet模型对U-net模型改进可以提高模型的评价指标,随着网络层数加深,评价指标有先增高,后趋于稳定的趋势,在网络层次达到101时评价指标达到最优,Res101-Unet模型的Accuracy、IoU、F1 Score分别为95.12%、60.13%、79.53%,对于简单和复杂裂隙的识别完整度都有提升;利用训练好的Res101-Unet模型对目标边坡上的裂隙进行识别,所得裂隙数量与现场测线方式所得结果一致,证明本模型识别结果与工程实际相符.

Joint fissures are one of the significant factors that influence the stability of open-pit mine slopes.With advancements in image processing and machine vision technology,the applications of intelligent algorithms for identification have attracted significant attention.Therefore,this paper proposes a method for identifying slope fissures in open-pit mines and deciphering geometric parameters,modernizing the U-net backbone network using residual network(ResNet)series algorithms for fast acquisition of joint fissure geometric information.The high-resolution images of open-pit mine slope fissures are collected using drones by considering factors such as viewpoint,distance,overlap rate,and flight speed.The images are subjected to preprocessing using the global threshold segmentation technique,and data augmentation is performed via random rotation,brightness,and contrast adjustment.The fissure image dataset then undergoes operations such as grayscale,threshold segmentation,dilation,hole filling,and the removal of small connected domain areas to eliminate the influence of background noise.Then,the U-Net network backbone is improved using five types of ResNet models:ResNet 18,34,50,101,and 152.This led to the proposed slope fissure recognition model based on the improved U-net network,which uses the pixel binary classification problem’s accuracy,Intersection over Union(IoU),and F1 Score as evaluation indicators.In addition,the proposed model is trained and assessed using the fissure image dataset.The fissure binary image output is compared with that of traditional fissure recognition methods.The Res101-Unet algorithm achieved accuracy(Pa)and IoU of 96.23%and 62.13%,respectively,offering finer and more extensive fissure recognition results than other methods.Geometric parameter information,such as fissure length and width distribution rules and parameters,is calculated from the fissure binary image.The results show an improvement in the model evaluation indicators owing to the enhancement of the INet model by the ResNet model.Furthermore,the accuracy of the index evaluation increases with the depth of the network layers.The Res101-Unet model reached its highest evaluation index when the number of network layers reached 101,with accuracy,IoU,and F1 scores reaching 95.12%,60.13%,and 79.53%,respectively.This scenario significantly improves the recognition of simple and complex fissures.As network layers deepen,fissure features can be captured from higher dimensions without substantially increasing network parameters.Thus,comprehensive and structurally distinct fissures can be obtained.The trained Res101-Unet model achieves the highest evaluation index upon reaching 101 network layers.Moreover,the number of recognized fissures on the target slope is consistent with the results obtained using the on-field measuring line method,confirming that the recognition results of this model are consistent with the actual engineering data.