文章摘要
翟晴* **,顾广华* **,孙雅倩* **,任贤龙* **.基于级联网络和语义层次结构的图像自动标注方法[J].高技术通讯(中文),2021,31(1):84~92
基于级联网络和语义层次结构的图像自动标注方法
  
DOI:10.3772/j.issn.1002-0470.2021.01.009
中文关键词: 图像自动标注; 级联网络; 行列式点过程(DPP); 语义层次结构(SH); 语义指标
英文关键词: automatic image annotation, cascade network, determinantal point process (DPP), semantic hierarchy (SH), semantic metrics
基金项目:
作者单位
翟晴* **  
顾广华* **  
孙雅倩* **  
任贤龙* **  
摘要点击次数: 1220
全文下载次数: 878
中文摘要:
      针对大多数的图像自动标注结果中含有冗余标签、信息量不够丰富的问题,本文提出了一种基于级联网络和语义层次结构的图像自动标注方法(CNSH)。首先,输入数据集的图片和标签列表,采用级联的VGG网络提取图像特征,训练条件行列式点过程(DPP)算法模型,计算标签的质量分数确定候选标签列表;其次,利用WordNet检索数据集标签得到语义层次结构和同义词,进而构建加权语义路径;最后,利用DPP算法在候选标签集中采样,得到最终的标注结果。与传统的图像标注任务相比,本文方法得到的标注结果能准确描述图片内容,且不含冗余标签。许多评估指标已用于图像标注和多标签学习,但是它们只专注于评估代表性,忽略了多样性。为了解决上述问题,本文采用了基于语义层次结构的语义指标来共同评估代表性和多样性。在IAPRTC-12和ESP Game 2个基准数据集上的实验表明,与现有方法相比本文方法能够产生更具代表性和多样性的标签。
英文摘要:
      There are some problems in automatic image annotation, such as the redundant labels and the lack of sufficient information. Aiming at the problem, an automatic image annotation method based on cascade network and semantic hierarchy (CNSH) is proposed. Firstly, from the input images and label list of dataset, image features are extracted through a cascaded VGG network. The condition determinantal point process (DPP) model is trained to compute the quality score of labels that is used to determine the list of candidate labels. Secondly, the semantic hierarchy and synonyms are obtained via a label set, WordNet to build a weighted semantic path. Finally, this work samples the candidate label set by using the DPP algorithm to obtain the final annotation results. Compared with the traditional image annotation methods, the annotation results are able to accurately describe the image content without redundant labels. Though many evaluation indexes are applied for the image annotation and the multi-label learning, they only focus on evaluating the representativeness and ignore the diversity. To solve the above drawbacks, the semantic metrics based on semantic hierarchy structure are applied to jointly evaluate the representativeness and diversity in this paper. Experimental results on IAPRTC-12 and ESP Game benchmark datasets demonstrate that the proposed method produces the more representative and diversified labels than the existing methods.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮