当前位置: X-MOL 学术Pattern Anal. Applic. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Neighborhood and center difference-based-LBP for face recognition
Pattern Analysis and Applications ( IF 3.9 ) Pub Date : 2021-01-02 , DOI: 10.1007/s10044-020-00948-8
Shekhar Karanwal , Manoj Diwakar

This research paper introduces the novel local binary pattern (LBP) variant for face recognition (FR) called as neighborhood and center difference-based-LBP (NCDB-LBP). In NCDB-LBP, the 4 labeled function is proposed to capture the robust features from 3 × 3 pixel window. For each neighborhood position , 2 first-order derivatives are computed, first computed between the adjacent neighborhood and the current neighborhood and the second computed between the center pixel and the current neighborhood. Employing the proposed function between the 2 first-order derivatives (produced from each neighborhood position) eventually results in 4 labeled window. All 8 neighborhoods are then placed in the 1 × 8 pixel window from which the 4 different binary patterns are produced. This concept is performed in both anticlockwise (ac) and clockwise (c) directions, termed as NCDB-LBPac and NCDB-LBPc descriptors. After binary patterns are encoded for each pixel position, the 4 transformed images are produced from ac direction and 4 from the c direction. All the respective directional transformed images are then divided into 3 × 3 subregions for histogram extraction. The combined histograms from all the respective subregions are the entire feature size of the NCDB-LBPac and NCDB-LBPc descriptors. To reduce the feature size, PCA and FLDA are utilized. Finally, classification is performed by SVMs and NN. The proposed FR approach is tested on ORL, GT, JAFFE, Yale, YB and EYB databases. The proposed FR approach achieves encouraging results.



中文翻译:

基于邻域和中心差异的LBP人脸识别

本研究论文介绍了一种新的用于人脸识别(FR)的本地二进制模式(LBP)变体,称为基于邻域和中心差的LBP(NCDB-LBP)。在NCDB-LBP中,提出了4标记函数来捕获3×3像素窗口中的鲁棒特征。对于每个邻域位置,计算2个一阶导数,第一个一阶导数在相邻邻域和当前邻域之间计算,第二个在中心像素和当前邻域之间计算。在2个一阶导数(从每个邻域位置生成)之间采用建议的功能最终会产生4个标记的窗口。然后将所有8个邻域放置在1×8像素窗口中,从中生成4个不同的二进制模式。这个概念是在逆时针(ac)和顺时针(c)方向上执行的,ac和NCDB-LBP c描述符。在为每个像素位置编码二进制模式之后,从ac方向生成4个变换图像,从c方向生成4个变换图像。然后将所有各自的定向变换图像划分为3×3个子区域,以进行直方图提取。来自所有各个子区域的组合直方图是NCDB-LBP ac和NCDB-LBP c描述符的整个特征尺寸。为了减小特征尺寸,使用了PCA和FLDA。最后,通过SVM和NN进行分类。在ORL,GT,JAFFE,Yale,YB和EYB数据库上测试了建议的FR方法。拟议的阻燃方法取得了令人鼓舞的结果。

更新日期:2021-01-02
down
wechat
bug