台标识别是典型的细微目标识别问题,针对台标区域小、信息量低,且镂空、半透明台标极易受到画面背景影响的难题,提出一个基于端到端全卷积网络的像素级台标识别网络——PNET.首先构建一个像素级标注的台标数据集,通过视频抽帧和图像预...台标识别是典型的细微目标识别问题,针对台标区域小、信息量低,且镂空、半透明台标极易受到画面背景影响的难题,提出一个基于端到端全卷积网络的像素级台标识别网络——PNET.首先构建一个像素级标注的台标数据集,通过视频抽帧和图像预处理获得台标图像集,并提出一种逐图像的像素级半自动标注方法获得二值标签图像集;然后提出一个像素级台标识别网络,在典型分类网络AlexNet,VGG的基础上,通过微调,将分类网络在分类任务中学习到的网络参数转换为像素级台标识别网络在台标分割任务中的所需的网络参数;最后引入跨层架构,融合来自网络深层的全局信息和浅层的局部信息.实验结果表明PNET实现了准确的像素级分割,准确率高达98.3%,在NVIDIA Tesla K80上单幅图像识别时间不超过1.5 s.展开更多
This research paper describes an SEE （Structural Engineering Encounter） Lab project. The paper reports on the development of a single-story, single-bay portal frame model as part of the AIMS2 （attract, inspire, men...This research paper describes an SEE （Structural Engineering Encounter） Lab project. The paper reports on the development of a single-story, single-bay portal frame model as part of the AIMS2 （attract, inspire, mentor and support students） grant supported through the US DOE （Department of Education） summer research program at California State University, Northridge. This research effort is part of a comprehensive program to develop laboratory models of structures commonly encountered in civil engineering practice, which can serve the dual purpose of accomplishing engineering education and research in the areas of structural and earthquake engineering. The objective of the present study was to construct a physical model of the aforementioned frame to experimentally collect data due to the application of vertical and lateral loadings through instrumentation such as strain gages and an LVDT （linear variable differential transformer） displacement transducer, and also to make comparisons with theoretical and numerical predictions.展开更多
基金863 program(No.2012AA011301)973 program(No.2010CB328204)+4 种基金NSFC project(Nos.61271189,61201154,60932004)RFDP Project(No.20120005120019)the Fundamental Research Funds for the Central Universities(No.2013RC1201)BUPT Excellent Ph.D.Students Foundation(No.CX201332)Fund of State Key Laboratory of Information Photonics and Optical Communications(BUPT)
文摘台标识别是典型的细微目标识别问题,针对台标区域小、信息量低,且镂空、半透明台标极易受到画面背景影响的难题,提出一个基于端到端全卷积网络的像素级台标识别网络——PNET.首先构建一个像素级标注的台标数据集,通过视频抽帧和图像预处理获得台标图像集,并提出一种逐图像的像素级半自动标注方法获得二值标签图像集;然后提出一个像素级台标识别网络,在典型分类网络AlexNet,VGG的基础上,通过微调,将分类网络在分类任务中学习到的网络参数转换为像素级台标识别网络在台标分割任务中的所需的网络参数;最后引入跨层架构,融合来自网络深层的全局信息和浅层的局部信息.实验结果表明PNET实现了准确的像素级分割,准确率高达98.3%,在NVIDIA Tesla K80上单幅图像识别时间不超过1.5 s.
文摘This research paper describes an SEE （Structural Engineering Encounter） Lab project. The paper reports on the development of a single-story, single-bay portal frame model as part of the AIMS2 （attract, inspire, mentor and support students） grant supported through the US DOE （Department of Education） summer research program at California State University, Northridge. This research effort is part of a comprehensive program to develop laboratory models of structures commonly encountered in civil engineering practice, which can serve the dual purpose of accomplishing engineering education and research in the areas of structural and earthquake engineering. The objective of the present study was to construct a physical model of the aforementioned frame to experimentally collect data due to the application of vertical and lateral loadings through instrumentation such as strain gages and an LVDT （linear variable differential transformer） displacement transducer, and also to make comparisons with theoretical and numerical predictions.
