【论文学习】Face Verification Using Deep Belief Networks

Icpr2016

Dong-Han Jhuang, Daw-Tung Lin∗, and Chi-Hung Tsai——National Taipei University

 

任务：

识别出一张人脸，有没有这个人脸

 

三步骤：

1、 feature extraction phase：

 

        range filtering：提取ROI，如把人脸从图像中抠出来

        down sampling：选部分的点

        Normal Vector Estimation：找脸表面每个点的法向量

        Principal Curvatures Estimation：使用PCA根据曲率求每个点的特征值（网络的输入）

2、 training phase：

网络：deep belief networks

结构：one visible layer comprising n neurons, two hidden layers having n/2 neurons, and one regression layer comprising two output labels

3、verification phase

 

实验结果

One thousand scans of 3D frontal facial point cloud  for each individual.

At the training stage, 500 point cloud scans were used as positive training samples and 500 samples were used as negative training examples（点云数据中没有人脸） for each individual.

 

论文参考：Face Verification with Three-Dimensional Point Cloud by Using Deep Belief Networks