电子学报 ›› 2017, Vol. 45 ›› Issue (12): 2925-2935.DOI: 10.3969/j.issn.0372-2112.2017.12.014

• 学术论文 • 上一篇    下一篇

基于蒙特卡罗数据集均衡与鲁棒性增量极限学习机的图像自动标注

柯逍1,2, 邹嘉伟1,2, 杜明智1,2, 周铭柯1,2   

  1. 1. 福州大学数学与计算机科学学院, 福建福州 350116;
    2. 福建省网络计算与智能信息处理重点实验室(福州大学), 福建福州 350116
  • 收稿日期:2016-03-14 修回日期:2017-03-27 出版日期:2017-12-25 发布日期:2017-12-25
  • 作者简介:柯逍,男,1983年10月生,福建省福州市人,博士,福州大学副教授,主要研究方向为计算机视觉、模式识别.E-mail:kex@fzu.edu.cn;邹嘉伟,男,1991年7月生,福建省龙岩市人,硕士,主要研究方向为计算机视觉、模式识别.E-mail:1468645610@qq.com;杜明智,男,1988年2月生,福建省泉州市人,硕士,主要研究方向为机器学习,图像处理.E-mail:dmz1028@163.com;周铭柯,男,1990年1月生,福建省三明市人,硕士,主要研究方向为深度学习、计算机视觉.E-mail:443810956@qq.com.
  • 基金资助:
    国家自然科学基金(No.61502105);福建省科技引导性项目(No.2017H0015);福建省中青年教师教育科研项目(No.JA15075)

The Automatic Image Annotation Based on Monte-Carlo Data Set Balance and Robustness Incremental Extreme Learning Machine

KE Xiao1,2, ZOU Jia-wei1,2, DU Ming-zhi1,2, ZHOU Ming-ke1,2   

  1. 1. College of Mathematics and Computer Science, Fuzhou University, Fuzhou, Fujian 350116, China;
    2. Fujian Provincial Key Laboratory of Networking Computing and Intelligent Information Processing, Fuzhou University, Fuzhou, Fujian 350116, China
  • Received:2016-03-14 Revised:2017-03-27 Online:2017-12-25 Published:2017-12-25

摘要: 针对传统图像标注模型存在着训练时间长、对低频词汇敏感等问题,该文提出了基于蒙特卡罗数据集均衡和鲁棒性增量极限学习机的图像自动标注模型.该模型首先对公共图像库的训练集数据进行图像自动分割,选择分割后相应的种子标注词,并通过提出的基于综合距离的图像特征匹配算法进行自动匹配以形成不同类别的训练集.针对公共数据库中不同标注词的数据规模相差较大,提出了蒙特卡罗数据集均衡算法使得各个标注词间的数据规模大体一致.然后针对单一特征描述存在的不足,提出了多尺度特征融合算法对不同标注词图像进行有效的特征提取.最后针对传统极限学习机存在的隐层节点随机性和输入向量权重一致性的问题,提出了鲁棒性增量极限学习,提高了判别模型的准确性.通过在公共数据集上的实验结果表明:该模型可以在很短时间内实现图像的自动标注,对低频词汇具有较强的鲁棒性,并且在平均召回率、平均准确率、综合值等多项指标上均高于现流行的大多数图像自动标注模型.

关键词: 蒙特卡罗数据集均衡, 多尺度特征融合, 极限学习机, 图像自动标注

Abstract: Aiming at the problem that the traditional image annotation model has long training time,sensitive to low-frequency words and other issues,this paper proposes a new automatic image annotation method based on Monte-Carlo dataset balance and robustness incremental extreme learning machine. First of all,training images of the public image library are segmented into different areas by this model and corresponding seed markup words are selected after segmentation,the areas are matched automatically based on comprehensive distance algorithm and the different keywords represent different areas. Then,for the huge difference of different annotated words' sizes in the public database,the Monte Carlo data set equalization algorithm is proposed to make the data size of each annotated word much the same. And a multi-scale feature fusion algorithm is proposed to extract effective features from different annotated images. Finally,the robustness incremental limit learning is proposed to improve the accuracy of the discriminant model for the problems of the consistency of the hidden layer nodes and the input vector weights existing in the traditional limit learning machine. The experimental results show that:compared with traditional algorithms of image automatic annotation,the methods proposed in this paper can implement the automatic image annotation quickly,and it is robust to low frequency words,and it is higher than most popular models of automatic image annotation in terms of average recall rate,average accuracy rate,comprehensive value and so on.

Key words: Monte-Carlo data set balance, multi-scale feature fusion, extreme learning machine, automatic image annotation

中图分类号: