基于SURF特征的栈式自编码网络人脸对齐算法

人脸对齐是人脸识别系统中的一个核心部分,定位的准确性和定位速度直接影响到人脸识别的效果。人脸图像存在不同姿态、不同表情、不同光照条件等因素的影响,真实场景下的人脸对齐成为一个难题。本文提出了一种基于SURF特征的栈式自编码网络人脸对齐方法,首先通过粗糙定位网络找到近似人脸特征点,并提取局部的SURF特征,输入到局部细化网络,通过级联结构,进一步精确化人脸特征点的具体位置。最后,在人脸数据集AFLW和HELEN上与近几年的对齐方法进行对比实验,平均错误率8.80%,i5四核CPU,2.3Hz主频硬件平台下计算时间7.6ms。我们的人脸对齐方法在真实场景下(包括单人和多人)具有较好的鲁棒性,可以实现准确定位。

Stacked auto-encoder networks face alignment algorithm based on SURF features

Face alignment is an important part of face recognition system. The accuracy and speed of alignment directly affect the effect of face recognition. Face images have different postures, different expressions, different lighting conditions and other factors. Face alignment becomes a difficult problem. This paper proposes a auto-encoder network face alignment algorithm based on SURF features. The approximate face feature points are found by the rough location network. The local SURF features are extracted and input into the local thinning network. The exact locations of the feature points of the face are further refined by the cascade structure. In the face data sets AFLW and HELEN, we compared some new alignment method. The average error rate is 8.80%, the computation time is 7.6 ms under i5 quad core CPU, 2.3 Hz dominant frequency. Our face alignment method has good robustness in real scene (including single and multiple faces), and it can achieve accurate positioning.