基于Voronoi K阶邻近图的半监督学习自动图像标注  

SEMI-SUPERVISED LEARNING FOR AUTOMATIC IMAGE ANNOTATION BASED ON VORONOI K-ORDER PROXIMITY GRAPH

作　　者：吴寿昆[1] 郭玉堂[2] 

机构地区：[1]安徽大学计算机科学与技术学院,安徽合肥230601 [2]合肥师范学院计算机学院,安徽合肥230601

出　　处：《计算机应用与软件》2016年第12期183-187,242,共6页Computer Applications and Software

摘　　要：在为自动图像标注构建相似图的过程中,针对传统的方法是基于图像间的视觉相似性,其没有考虑到数据集中某个子数据集内的结构信息这一问题,提出一种基于Voronoi k阶邻近图的半监督学习自动图像标注方法。该方法充分考虑Voronoi k阶邻近图能很好地表达空间目标的影响区域以及可以方便地进行空间邻近的描述与推理的特性,将特征空间内的图像数据点分布信息融合到点对间的相似度量表示中,利用未标注样本挖掘图像特征的内在规律,然后把半监督学习的方法和多标记学习有效结合起来,从而达到对图像进行自动标注。实验结果表明,提出的标注方法可行,同时标注结果与传统的标注方法相比得到了明显改善。The traditional method in the process of constructing a similar graph for automatic image annotation is based on the visual similarity between images, ignoring the structural information of the data set, thus a semi-supervised learning method for automatic image annotation method for is proposed based on Voronoi k-order proximity graph through the summary and analysis of existing research results. The Voronoi k- order proximity graph is able to express the impact area of the target space well and describe and infer the spatial proximity conveniently, which is fully used by this method. Then the distribution information of the data points in the feature space is integrated into the similarity rep- resentation in two nodes. The method combines the method of semi-supervised learning with multi-label learning by using the inherent law of digging out image feature through unlabeled samples so as to achieve automatic image annotation. The experimental results show that the proposed method is feasible, and the annotation results are improved compared with the traditional method.

关 键 词：半监督学习 VORONOI k阶邻近图 自动图像标注 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]