摘要
为实现深度学习图像处理技术快速获取露天矿山路网信息,搭建了Deeplabv3+与PSPNet 2种典型网络模型框架,选取不同层数的ResNet为骨干网络构建起露天矿道路信息提取模型,通过框架模型横向对比和同框架下不同层数的ResNet纵向对比,研究了模型识别精度和速度与ResNet层数间的最优分配问题。结果表明:综合考虑模型训练时长与识别效果的前提下,Deeplabv3+网络框架对于图像边缘区域的识别存在道路信息丢失、无法识别或误差较大的问题,不宜应用于露天矿山道路信息的图像处理;PSPNet网络框架能够准确识别复杂矿区环境中的多类不同特征的道路,且骨干网络ResNet的层数保持在50层即可满足识别精度的同时保证识别速度与效率。
In order to achieve rapid acquisition of open-pit mine road network information by deep learning image processing technology,the article constructs two typical network model frameworks,Deeplabv3+and PSPNet,selects ResNet with different layers as the backbone network to construct an open-pit mine road information extraction model.Through horizontal comparison of the framework model and vertical comparison of ResNet with different layers under the same framework,the article studies the optimal allocation problem between model recognition accuracy and speed and ResNet layers.The results show that taking into account both the training time and recognition performance of the model,the Deeplabv3+network framework has problems with road information loss,inability to recognize,or large errors in recognizing image edge areas,and is not suitable for image processing of road information in open-pit mines.The PSPNet network framework can accurately identify multiple types of roads with different features in complex mining environments,and the backbone network ResNet can maintain a layer count of 50 to meet the recognition accuracy while ensuring recognition speed and efficiency.
作者
田培忠
TIAN Peizhong(Ha’erwusu Open-pit Mine,China Energy Group Zhunge’er Energy Co.,Ltd.,Ordos 010300,China)
出处
《露天采矿技术》
CAS
2024年第5期59-64,共6页
Opencast Mining Technology
