计算机与现代化

• 图像处理 • 上一篇    下一篇

基于局部相位纹理表示的光照变化人脸识别算法

  

  1. 无锡太湖学院工学院,江苏无锡214064
  • 收稿日期:2015-07-24 出版日期:2015-12-23 发布日期:2015-12-30
  • 作者简介:王华君(1979-),男,江苏宜兴人,无锡太湖学院工学院讲师,硕士,研究方向:图像处理,模式识别; 李荣(1978-),女,江苏淮安人,讲师,硕士,研究方向:图像处理,人工智能; 徐燕华 (1979-),女,江苏无锡人,讲师,硕士,研究方向:图像处理,智能算法; 孟德建(1979-),男,江苏扬州人,讲师,博士,研究方向:图像处理,视频处理。
  • 基金资助:
    江苏省高校自然科学研究项目(14KJB520036)

A Face Recognition Algorithm Under Illumination Variation Based on #br# Local Phasetexture Representation

  1. School of Engineering, Taihu University of Wuxi, Wuxi 214064, China
  • Received:2015-07-24 Online:2015-12-23 Published:2015-12-30

摘要:

针对基于局部纹理的人脸表示不能较好解决不同光照条件下低分辨率人脸图像识别的问题,提出一种新的相位纹理表示法。该算法在局部邻域使用傅里叶变换相位的四象限掩码,减少来自更高幅
值响应的滤波器中的错误滤波响应影响,从而产生更具判别性的代码滤波响应,相比局部相位量化(Local Phase Quantization, LPQ)受噪声影响大、量化离散效应等影响,相位纹理表示法更加有效和稳
定。在CMUPIE、扩展YALEB和AR人脸数据库上的实验结果表明,本算法比局部相位量化更具描述性,识别率比LPQ和广泛使用的LBP(Local Binary Pattern, LBP)和方向梯度直方图(Histogram of
Oriented Gradients, HOG)均有较大幅度的提高,对于增强光照条件,识别率增益小于1%,对光照变化的鲁棒性优于其他3种算法。

关键词: 光照变化, 人脸识别, 相位纹理表示, 傅里叶变换, 局部相位量化, 滤波响应

Abstract:

As the algorithms based on local texture for face recognition can not well solve the problem of low resolution face recognition under varying illumination conditions,
a new phase texture representation is presented. The basic idea is to use the fourquadrant phase mask of Fourier Transform in a local neighborhood, with reducing the amplitude
response of the filter from the higher impact of the error filter response. And this can generate coding filter response with more discriminating. Comparing with local phase
quantization (LPQ) that is affected by noise impact and the impact of discrete effects, phase texture representation is more effective and stable. The experimental results on
the three databases CMUPIE, extended YALEB and AR show that the proposed algorithm is more descriptive than LPQ recognition and the widely used description like local binary
pattern (LBP), histogram of oriented gradients (HOG). For enhanced lighting conditions, the recognition grain rate of proposed algorithm is less than 1 percent, much better than
the three other algorithms in robust to illumination changes.

Key words: illumination variation, face recognition, phase texture representation, Fourier transform, local phase quantization, filter response

中图分类号: