基于连续预测的半监督学习图像语义标注  

Semi-supervised learning image semantic annotation based on sequential prediction

作　　者：郭玉堂[1,2] 李艳[1] 

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科学与技术系,安徽合肥230601

出　　处：《计算机工程与科学》2015年第3期553-558,共6页Computer Engineering & Science

基　　金：安徽省自然科学基金资助项目(11040606M134);安徽省高校自然科学基金资助项目(KJ2103A217)

摘　　要：为了在图像底层特征与高层语义之间建立关系,提高图像自动标注的精确度,结合基于图学习的方法和基于分类的标注算法,提出了基于连续预测的半监督学习图像语义标注的方法,并对该方法的复杂度进行分析。该方法利用标签数据提供的信息和标签事例与无标签事例之间的关系,根据邻接点(事例)属于同一个类的事实,构建K邻近图。用一个基于图的分类器,通过核函数有效地计算邻接信息。在建立图的基础上,把经过划分后的样本节点集通过基于连续预测的多标签半监督学习方法进行标签传递。实验表明,提出的算法在图像标注中的标注词的平均查准率、平均查全率方面有显著的提高。In order to establish the relationship between low-level features and high-level semantics of the image,improve the accuracy of image automatic annotation,combining with graph learning and classification annotation algorithm,we propose an image semantic annotation method for sequential predictionbased semi-supervised learning,and analyze the complexity of the method.According to the fact that the adjacent vertexes（cases）should belong to the same class,by using the information provided by tag datum and the relationship between tag cases and cases with no labels,the method constructs a K relative neighborhood graph.We use a graph-based classifier and a kernel function to calculate the adjacency information effectively.On the basis of building graphs,we propagate the labels of the node sets derived from the samples by sequential prediction-based semi-supervised multiple labels learning method.Experiments show that the proposed algorithm for image annotation significantly improves the average precision ratio and the average recall ratio of the marked words.

关 键 词：连续预测 半监督 图像标注 图学习 多标签 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]