基于语音识别的蔬菜病害视频语义标注与分割方法  

Video Semantic Annotation and Segmentation Method of Vegetable Disease Knowledge Based on Voice Recognition

作　　者：李鑫星[1] 刘春迪[1,2] 温皓杰[1,2] 苏叶[1,2] 傅泽田[1,3] 张领先[1,2] 

机构地区：[1]中国农业大学信息与电气工程学院,北京100083 [2]农业部农业信息获取技术重点实验室,北京100083 [3]食品质量与安全北京实验室,北京100083

出　　处：《农业机械学报》2015年第9期308-313,共6页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家自然科学基金资助项目(31271618);现代农业产业技术体系北京市叶类蔬菜创新团队建设专项科研资金资助项目(blvt-20);中央高校基本科研业务费专项资金资助项目(2013XJ021);北京市大学生科学研究与创业行动计划资助项目(014bj091)

摘　　要：为了向农民提供蔬菜病害知识,基于语音识别技术设计了一种蔬菜病害视频标注与分割方法,可将科研机构录制的蔬菜病害视频分割成适合手机播放的小视频段落。在前期设计的视频镜头切分方法基础上,进一步设计出基于语音识别技术的视频语义标注及视频镜头聚类方法,即首先采用成熟的语音识别技术,将视频镜头的语音讲解识别为文本形式;进而基于本体对识别文本进行相应的语义处理,从中提取出能起到指示作用的关键语义实体,并将其恰当的组织形式作为视频镜头的语义标注;最终根据用户提供的关键词并结合视频镜头的语义标注,对视频镜头进行聚类和重组,从而实现对于蔬菜病害视频的最终分割。所设计的视频镜头语义标注方法对2个测试视频的查全率分别达到96.08%、94.93%,查准率分别达到94.31%、95.98%,F-1测度也分别达到0.93和0.92;视频镜头聚类方法使得2个视频的分割查全率分别达到94.9%、98.7%,查准率分别达到92.1%、90.2%,查全率平均大于95%,查准率大于90%。证明所设计的蔬菜病害视频标注与分割方法具有理论和实用价值。To provide farmers with vegetable diseases knowledge, this paper proposes a method based on voice recognition technology to label and split vegetable diseases videos. Through this method, videos about vegetable diseases can be split into several smaller segments which are more suitable for cell phone. The methods of semantic annotation and video shot clustering were based on video segmentation and voice recognition. In this method, the audio signals of videos were transformed into text strings firstly by voice recognition. Then key semantic entities for labelling video shots semantically were split from the text strings. Finally different video shots were clustered and recombined based on keywords provided by user arid the semantic labels of video shots. When applying the method of semantic annotation to two videos, the recall ratios were up to 96.08% and 94. 93% , the precision ratios were up to 94. 31% and 95.98%, and the F- 1 measures were up to 0. 93 and 0. 92. As for method of video shot clustering, the recall ratios were up to 94.9% and 98.7%, and the precision ratios were up to 92. 1% and 90. 2%. Results of comparative experiments show that the proposed method is valuable both in theory and practice.

关 键 词：蔬菜病害 视频分割 语音识别 语义标注 

分 类 号：S2[农业科学—农业工程]