作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2011, Vol. 37 ›› Issue (8): 166-168. doi: 10.3969/j.issn.1000-3428.2011.08.057

• 人工智能及识别技术 • 上一篇    下一篇

基于多尺度LBP金字塔特征的分类算法

梁 鹏 1,何俊诗 2,黎绍发1   

  1. (1. 华南理工大学计算机科学与工程学院,广州 510006;2. 广州市番禺中心医院,广州 511400)
  • 出版日期:2011-04-20 发布日期:2012-10-31
  • 作者简介:梁 鹏(1981-),男,博士研究生,主研方向:计算机视觉,模式识别;何俊诗,在职硕士研究生;黎绍发,教授、博士生导师
  • 基金资助:
    国家自然科学基金资助项目(60273064);广东省工业攻关计划基金资助项目(2004B10101032)

Classification Algorithm Based on Multi-scales LBP Pyramid Feature

LIANG Peng 1, HE Jun-shi 2, LI Shao-fa 1   

  1. (1. School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, China; 2. Guangzhou Panyu Centeral Hospital, Guangzhou 511400, China)
  • Online:2011-04-20 Published:2012-10-31

摘要: 为有效解决旋转变化、光照变化和尺度变化等图像的分类问题,提出一种基于多尺度局部二元模式(LBP)金字塔特征的图像分类算法。通过多尺度LBP金字塔提取各尺度的图像纹理特征,建立图像的多尺度LBP金字塔直方图,并将其作为图像特征向量,采用K-means方法对该特征向量进行降维,以用于图像分类。同时,针对传统二进制权值分布方法对噪声敏感的缺点,提出一种多端权值分布方法。实验结果表明,多尺度LBP金字塔方法具有较好的可鉴别性及图像描述能力,而多端权值分布法也能提高图像的分类精度。

关键词: 多尺度LBP金字塔, 图像分类, 多端权值分配, 特征向量

Abstract: In order to effectively solve the rotation changes, lighting changes and scale changes in the problem of image classification difficulties, a novel image classification algorithm based on multi-scales Local Binary Pattern(LBP) pyramid feature is proposed. It builds a multi-scales LBP pyramid to extract image texture features. It makes up the multi-scales LBP pyramid histograms by the extracted features. The dimensionality of multi-scales LBP pyramid histograms is reduced by kmeans clustering for image classification. The analysis of experimental results proves that the proposed algorithm has better discriminative power and image description ability. Furthermore, aiming at the drawback of convention weighting scheme based on the binary feature matching, it also proposes a new weighting scheme named multi-dominant feature weighting. Experimental results show this method actually improves the performance of classification.

Key words: Multi-scales Local Binary Pattern(LBP) pyramid, image classification, multi-dominant weighting distribution, feature vector

中图分类号: