Page 109 - 2021年第52卷第9期
P. 109
[ 24] ZHOU Y,STARKEY J,MANSINHA L . Segmentation of petrographic images by integrating edge detection and
region growing[J]. Computer Geosciences,2004,30:817-831 .
[ 25] 商梦石 . 基于图像处理的矿石粒度检测方法研究[D]. 昆明:昆明理工大学,2017 .
[ 26] YANG W,WANG K,ZUO W . Neighborhood component feature selection for high-dimensional data[J]. JCP,
2012,7(1):161-168 .
[ 27] 李明超,符家科,张野,等 . 耦合岩石图像与锤击音频的岩性分类智能识别分析方法[J]. 岩石力学与工程
学报,2020,39(5):996-1004 .
[ 28] 周胜,刘三民 . 基于迁移学习的数据流分类研究综述[J]. 天津理工大学学报,2019,35(3):24-29 .
[ 29] 闫 涵 ,张 旭 秀 ,张 净 丹 . 多 感 知 兴 趣 区 域 特 征 融 合 的 图 像 识 别 方 法[J]. 智 能 系 统 学 报 ,2021,16(2):
263-270 .
[ 30] 司晨冉,王仁超,邸阔,等 . 一种基于 Mask R-CNN 和分水岭算法的岩石颗粒图像分割方法[J]. 水电能源
科学,2020,38(11):129-132,128 .
[ 31] 雷 雨 萌 ,陈 祖 煜 ,于 沭 ,等 . 基 于 深 度 阈 值 卷 积 模 型 的 土 石 料 级 配 智 能 检 测 方 法 研 究[J]. 水 利 学 报 ,
2021,52(3):369-380 .
[ 32] 杨亚男,张宏鸣,李杭昊,等 . 结合 FCN 和 DenseCRF 模型的无人机梯田识别方法研究[J]. 计算机工程与
应用,2021,57(3):222-230 .
[ 33] KRAHENBUHL P,KOLTUN V . Efficient inference in fully connected CRFs with Gaussian Edge Potentials[J].
Advances in Neural Information Processing Systems,2011,24:109-117 .
[ 34] CHENG S H,MA J Y,ZHANG S J . Smoke detection and trend prediction method based on Deeplabv3+ and gen⁃
erative adversarial network[J]. Journal of Electronic Imaging,2019,28(3):033006 . doi:10.1117/1.JEI.28.
3.033006.
[ 35] 谢 梦 ,刘 伟 ,李 二 珠 ,等 . 深 度 卷 积 神 经 网 络 支 持 下 的 遥 感 影 像 语 义 分 割[J]. 测 绘 通 报 , 2020(5):
36-42 .
[ 36] DL/T 5395-2007,碾压式土石坝设计规范[S]. 北京:中国电力出版社,2008 .
[ 37] SL 251-2015,水利水电工程天然建筑材料勘察规程[S]. 北京:中国水利水电出版社,2015 .
[ 38] GB/T 50123-2019,土工试验方法标准[S]. 北京:中国计划出版社,2019 .
[ 39] 王仁超,朱品光 . 基于随机森林回归方法的爆破块度预测模型研究[J]. 水力发电学报,2020,39(1):
89-101 .
Intelligent detection method of rockfill particle size distribution
based on Deep-learning and NCFS algorithm
1
1
WANG Renchao ,LIAN Jiaxin ,DI Kuo 1,2
(1. State Key Laboratory of Hydraulic Engineering Simulation and Safety,Tianjin University,Tianjin 300350,China;
2. Frontier Technology Research Institute,Tianjin University,Tianjin 301700,China)
Abstract:Focus on the problems such as the failure of rapid detection of blasting material particle size by
artificial screening,and the low accuracy and poor generalization of the existing particle size detection mod⁃
el in the construction process of rockfill dam, a method which combined Deep-learning and NCFS algo⁃
rithm to detect the particle size of rockfill dams in digital screening is presented,the method could detect
the particle size through the images of blasting pile quickly. Aiming at improving the accuracy of training
result of Deep-learning model, the DenseCRF algorithm is utilized to optimize the result of image training
by Deeplabv3+ model. In order to transform the 2D features extracted from image to 3D particle size distri⁃
bution of pile,the NCFS algorithm is exploited to characterize the relationship between 2D parameters and
3D particle size, and a software to realize the transformation is programmed by MATLAB. Through images
acquisition and screening test analysis of blasting piles in Jurong pumped storage power station, the result
shows that,compared with other methods,the proposed method is feasible,and the accuracy of feature ex⁃
traction and particle size prediction is improved.
Keywords:size detection of rockfill dams;deep-learning;Deeplabv3+;DenseCRF;NCFS algorithm
(责任编辑:耿庆斋)
— 1115 —