摘要
注意力检测依次分为三步:人脸五特征点提取,单应性矩阵计算面部转角以及判断注意力是否集中。人脸五特征点提取对最终的检测结果起决定性作用,目前基于深度学习的五特征点检测方法已经越来越占据主流。为了提高特征点检测的实时性,对通用的DCNN进行优化,将原来的三级预测改为两级。通过实验发现将三级网络改为两级网络在缩短时间的同时特征点的检测精度还有微小的提升,面部特征点的检测准确率达到98.11%。之后将新网络检测出的五特征点应用到人脸偏转角度计算中,得到人脸偏转的角度误差在5度以内,单应性矩阵计算出的鼻子像素点准确率高达98.99%。
Attention detection is divided into three steps: extraction of five facial feature points, calculation of face angles by homography matrix and judgment of attention. Five facial feature point extraction plays a decisive role in the final detection results. At present, five feature point detection methods based on deep learning have become more and more mainstream. In order to improve the real-time performance of feature point detection, the general DCNN was optimized, and the original three-level prediction was changed into a two-level prediction. Through the experiment, we found that the change of a three-level network to a two-level network not only shortened the time, but also slightly improved the detection accuracy of feature points, and the detection accuracy of facial feature points reached 98.11%. Then, the five feature points detected by the new network were applied to the face deflection angle calculation, and the error of face deflection angle was less than 5 degrees. The accuracy of nose pixel points calculated by the homography matrix was 98.99%.
作者
郭克友
贺云霞
胡巍
GUO Ke-you;HE Yun-xia;HU Wei(Beijing Technology and Business University,Beijing 100048,China)
出处
《计算机仿真》
北大核心
2022年第12期240-244,共5页
Computer Simulation
关键词
实时
卷积神经网络
特征点检测
单应性矩阵
人脸角度偏转
Real-time
Convolutional neural network
Feature point detection
Homography matrix
Face angle deflection