CN201810023920
CN108052680A
微信扫码查看/分享专利
|
摘要:本发明是一种基于数据图谱、信息图谱、知识图谱架构的图像数据目标识别增强方法。主要用于解决现有图像识别方法无法识别训练集中未标注类别的图像识别问题,属于分布式计算和软件工程学技术交叉领域。关键在于从已有的基于深度学习方法的图像类型识别结果出发,根据已有图像资源构建三层图谱,将未识别的图像类别在数据图谱进行特征匹配得到初始匹配结果,将已识别的图像类别在信息图谱上进行关系匹配得到中间匹配结果,最后在知识图谱中进行间接交互关系匹配,计算中间匹配结果的可信度并排序,推荐给用户可信度最高的匹配图像类别。
|
Abstract: The invention discloses an image data target recognition enhancing method based on a data map, an information map and a knowledge map, mainly aims to solve the problem that existing image recognitionmethods cannot recognize, train and centralize images with types unmarked and belongs to the technical cross field of distributed computing and software engineering. The method includes: starting froman existing image type recognition result based on an in-depth learning method, establishing a three-layer map according to known image resources; subjecting unrecognized image type to feature matching on the data map to obtain an initial matching result; subjecting recognized image type to relation matching on the information map to obtain an intermediate matching result; performing indirect interaction relation matching in the knowledge map, calculating credibility of the intermediate matching result, ranking, and recommending the matching image type with highest credibility to a user.
|
申请人: 海南大学
|
Applicant: UNIV HAINAN
|
地址: 570228 海南省海口市美兰区********(隐藏)
|
发明(设计)人: 段玉聪 何诗情 靖蓉琦 宋正阳 邵礼旭
|
Inventor: SHIQING HE; YUCONG DUAN; ZHENGYANG SONG; LIXU SHAO; RONGQI JING
|
主分类号: G06F17/30(2006.01)I
|
分类号: G06F17/30(2006.01)I
|
|