摘要
针对基于深度学习的人脸识别算法模型的单一尺度输入问题,以及人脸图像样本在输入模型过程中由尺寸放缩而导致人脸特征信息丢失问题,提出一种基于深度学习的多尺度轻量化的人脸识别算法。同时,鉴于现有人脸数据集的同一样本单一尺度的局限性,提出一种多尺度的人脸数据集。该算法首先建立一种单一尺度的人脸识别模型,然后通过空间金字塔池化结构实现模型的多尺度输入。再选用Maxout作为模型的激活函数,实现模型的轻量化。数据集首先建立距摄像机不同距离下的三个采集点,然后在每个采集点上采集不同角度的人脸图像,最后对采集到的人脸图像进行数据预处理和分类整理。实验结果表明,所提出的算法在距离条件变化的情况下取得较好的性能。
Aiming at the single-scale input problem of face recognition algorithm model based on deep learning,and the loss problem of face features information caused by size scaling of face image samples in the process of inputting the model,proposes a multi-scale lightweight face recog-nition algorithm based on deep learning.At the same time,in view of the limitations of the same sample single scale of the existing face datas-et,proposes a multi-scale face dataset.The algorithm first establishes a single-scale face recognition model,and then realizes the multiscale input of the model through the structure of spatial pyramid pooling.Then uses Maxout as the activation function of the model to achieve lightweight of the model.The dataset first establishes three collection points at different distances from the camera,then collects different an-gles of face images at each collection point,and finally data preprocessing and labeling are performed on the collected face images.The ex-perimental results show that the proposed algorithm achieves better performance under the condition of changing distance conditions.
作者
张文涛
陈婵娟
王泽荔
ZHANG Wen-tao;CHEN Chan-juan;WANG Ze-li(College of Mechanical&Electrical Engineering,Shaanxi University of Science&Technology,Xi’an 710021)
出处
《现代计算机》
2018年第20期31-37,共7页
Modern Computer
关键词
人脸识别
卷积神经网络
轻量化模型
图像金字塔
Face Recognition
Convolution Neural Network
Lightweight Model
Image Pyramid
作者简介
张文涛,男,硕士,研究方向为机器视觉与模式识别;陈婵娟,女,教授,研究方向为机械电子工程;王泽荔,男,硕士,研究方向为机器视觉与模式识别。
