摘要
针对表情变化下的三维人脸识别问题,提出了一种基于几何图像滤波的特征提取方法,并根据样本图像滤波后的特征分布函数给出最优卷积滤波器的设计过程.首先,利用网格平面参数化方法,将人脸网格映射到边界为正四边形的平面区域内,经过线性插值采样得到具有三维形状的二维几何图像;然后,将整体几何图像切割成局部分块图像的集合,在每组局部分块图像构成的训练样本库中利用差分进化算法对滤波器进行优化设计;最后,利用训练得到的最优滤波器提取对应分块图像的局部特征并计算相似度,将相似度得分融合,即可得到最终识别结果.利用FRGC v2人脸数据库进行实验验证,结果表明,使用几何图像滤波能显著提高算法的精度和鲁棒性.
Aiming at the 3D face recognition under expression variation, a feature extraction method by applying filter on geometry image is proposed. The design for the optimal convolution filter is presented based on the distribution function of filtered features. First, after objectively mapping facial mesh into square domain based on mesh parameterization, a 2D geometry image with 3D shape is obtained by linear interpolation. Then, the entire images in the training set are segmented into pat- ches which are used for the differential evolution algorithm to design the optimal convolution filters. Finally, the similarity scores between local features are computed by applying these filters on corre- sponding patches, and the final decision is made by combining results of these scores. The experi- mental results from FRGC （face recognition grand challenge） v2 （version 2 ） databases show that both accuracy and robustness are improved by applying filter on geometry image.
出处
《东南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2012年第5期859-863,共5页
Journal of Southeast University：Natural Science Edition
基金
国家自然科学基金资助项目(51175081)
江苏省自然科学基金资助项目(BK2010058)
关键词
3D人脸识别
形状滤波
几何图像
卷积滤波
3D face recognition
shape filtering
geometry image
convolution filter
