期刊文献+

基于深度学习的部分人脸预测及识别 被引量:5

Partial face prediction and recognition based on deep learning
下载PDF
导出
摘要 左右划分后的人脸被称为纵向部分人脸。为了衡量纵向部分人脸的大小,在尽可能少损失识别精度的情况下充分利用纵向部分人脸,提出一种简单有效的解决方案。首先,将纵向部分人脸的大小按照其占整张人脸的宽度比量化。其次,采用深度学习方法设计一种人脸比例特征网络。通过训练大量的部分人脸数据,使用该神经网络预测纵向部分人脸的宽度比,即部分人脸的大小。最后,通过分析不同大小的纵向部分人脸对人脸验证的影响程度判别出最佳宽比0. 7并作为比例阈值,当人脸比例特征网络预测结果超过比例阈值0. 7时,可用作后续人脸识别,否则丢弃该部分人脸图像。占宽比预测实验结果显示,在所选10个点处,平均误差在实际宽比1. 0时达到最大0. 035 735,在0. 1处达到最小0. 006 857。误差百分比在7%以下,方差最高不超过0. 000 352。 Faces divided to the left and right are called vertical partial faces. In order to measure the size of a vertical partial face, and to make full use of the vertical part of the face with recognition precision loss as little as possible, a simple and effective solution was presented. Firstly, the size of the vertical partial face was quantified as the width ratio of the face. Secondly, deep learning was used to design a face-proportion-feature network. By training a large number of face data, a neural network was used to predict the width ratio of the vertical partical face, that is, the size of partial face. Finally, through the analysis of the influence of face width ratio on face verification, the optimal width ratio 0.7 was identified as the threshold. When the prediction result of face-proportion-feature network exceeded the threshold 0.7, the subsequent face recognition can be can'ied on, otherwise the face image was discarded. The results of the width ratio prediction experiment show that the average error reaches the maximum of 0. 035 735 at the actual width ratio of 1 at the selected ten points, and the minimum of 0. 006 857 at 0.1. The error is less than 7%, and the maximum of variance is no more than 0. 000 352. It shows that face- proportion-feature network can predict the width ratio of the vertical partial face very well.
作者 居聪 李晓峰 黄飞虎 JU Gong;LI Xiaofeng;HUANG Feihu(College of Computer Science,Sichuan University,Chengdu Sichuan 610065,China;National Key Laboratory of Fundamental Science on Synthetic Vision(Sichuan University),Chengdu Sichuan 610065,China)
出处 《计算机应用》 CSCD 北大核心 2018年第A01期36-38,103,共4页 journal of Computer Applications
基金 国家重大科学仪器设备开发专项(2013YQ490879)
关键词 人脸识别 纵向部分人脸 深度学习 宽度比 神经网络 face recognition vertical partial face deep learning width ratio neural network
  • 引文网络
  • 相关文献

参考文献2

二级参考文献7

共引文献20

同被引文献23

引证文献5

二级引证文献31

;
使用帮助 返回顶部