一种大规模图像多样化语义标注方法  

Diverse Image Annotation Method for Large Scale Data Set

作　　者：田枫[1] 沈旭昆[2] 周凯[1] 

机构地区：[1]东北石油大学计算机与信息技术学院,大庆163318 [2]北京航空航天大学虚拟现实技术与系统国家重点实验室,北京100191

出　　处：《系统仿真学报》2014年第9期2085-2090,共6页Journal of System Simulation

基　　金：国家高技术研究发展计划(2009AA012103);东北石油大学青年科学基金(2013NQ120;2012QN117);黑龙江省教育厅科学技术研究项目资助(12521055)

摘　　要：提出了一种大规模图像多样化语义标注方法,以解决传统的图像语义标注结果中存在的语义单一化问题。语义标注多样化要求图像的语义标签不仅要和图像内容相关,也要和图像的其他标签具有语义差别,从多个角度对图像内容进行描述。定义了一个融合相关性和多样性的度量。利用视觉语言模型分别计算标签与图像的相关性和标签之间的视觉距离。基于平均多样化准确率期望的最大化,通过启发式的迭代求解过程得到兼具相关性和多样性的标签集合。网络数据集上的实验结果表明,该方法在相关性保持的基础上,提高了主题覆盖率,涵盖了更广泛的语义概念。A diverse image annotation method was proposed, aiming to take relevance and diversity into account by exploring the content of images and their associated labels. The measure of average diverse precision was proposed to evaluate the diversity and precision of annotation result. The visual language model was employed to obtain the relevance between images and labels, as well as the visual distance between labels. The annotation result list was generated by a heuristic and iterative algorithm which optimizes the expected value of average diverse precision, which considered both relevance and diversity. Experiments on web dataset show that the proposed method obtained diverse results with relevance maintained.

关 键 词：图像标注 多样化标注 语义标注 大规模数据集 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]