PNET:像素级台标识别网络  

PNET: Pixel-wise TV Logo Recognition Network

作　　者：徐佳宇 张冬明 靳国庆[3] 包秀国 袁庆升 张勇东 Xu Jiayu;Zhang Dongming;Jin Guoqing;Bao Xiuguo;Yuan Qingsheng;Zhang Yongdong(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072;The National Computer Network Emergency Response Technical Team Coordination Center of China,Beijing 100029;Intelligent Information Processing Key Laboratory,Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190;School of Information Science and Technology,University of Science and Technology of China,Hefei 230026;School of Cyber Security,University of Chinese Academy of Sciences,Beijing 100193)

机构地区：[1]天津大学电气自动化与信息工程学院,天津300072 [2]国家计算机网络应急处理协调中心,北京100029 [3]中国科学院计算技术研究所智能信息处理重点实验室,北京100190 [4]中科院信息工程研究所中国科学大学网络空间安全学院,北京100193 [5]中国科学技术大学信息科学技术学院,合肥230026

出　　处：《计算机辅助设计与图形学学报》2018年第10期1878-1889,共12页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家重点研发计划(2016YFB0801203);国家自然科学基金(61672495,61273247).

摘　　要：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.TV logo recognition is a typical fine object recognition problem,referring to the problem that TV logo region is small and contains low amount of information,hollow-out and translucent logos are easily influenced by background in video frame,a pixel-wise TV logo recognition network based on an end to end fully convolutional network was proposed.Firstly a pixel-wise annotated TV logo dataset was constructed,a TV logo image set was obtained by extracting and preprocessing video frames,and a binary label image set was obtained by proposing a pixel-wise semi-automatic annotation method.Then a pixel-wise TV logo recognition network PNET was proposed based on a typical classification network AlexNet or VGG,and network parameters learned by a classification network in a classification task were converted to the network parameters required by a pixel-wise TV logo recognition network in a segmentation task.Finally a skip architecture was introduced in network combining global information from deep layers and local information from shallow layers.The experiment results show that PNET achieves accurate pixel-wise segmentation.The accuracy is up to 98.3%and inference time for per image on NVIDIA Tesla K80 is less than 1.5 s.

关 键 词：视频分类 台标识别 全卷积网络 像素级半自动标注 跨层架构 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]