摘要
由于人脸姿态、表情、遮挡物、光照问题的影响,人脸关键点检测时通常会出现较大的误差,为了准确且可靠地检测关键点,提出了一种基于级联卷积神经网络的方法。利用人脸检测器检测到的人脸图像作为输入,第一层卷积神经网络直接检测所有的5个人脸关键点。随后根据这些检测到的点裁剪出5个人脸局部图像,级联的第二层网络使用5个不同的卷积神经网络单独地定位每个点。在实验测试环节,级联卷积神经网络方法的使用将人脸关键点的平均定位误差降低到了1.264像素。在LFPW人脸数据库上的实验结果表明:该算法在定位准确性和可靠性上要优于单个CNN的方法以及其他方法,该算法在GPU(图形处理器)模式下处理一个人脸图像仅需15.9毫秒。
Suffered from facial pose, expressions, occlusions and illumination, there is usuallylager errorsin thefacial points detection. In order to detect facial points accurately and reliably, a method based on cascade convolutional network is proposed. Using the face image detected from face detector as input, all of 5 facial points are detected bythe first level CNN directly. After then 5 images from the facial points are cropped, and each of the points is predicted singly by the second level CNN with 5 different CNN. The mean error of all points is reduced to 1. 264 pixel in test phase by the cascade convolutional network method. The experiments on LFPW database show that this method outperforms single CNN method and many other methods in both detection accuracy and reliability. This method process one face image takes approximately 15.9 ms on a standard GPU （graphics processing unit）.
出处
《四川理工学院学报（自然科学版）》
CAS
2017年第1期32-37,共6页
Journal of Sichuan University of Science & Engineering(Natural Science Edition)
