«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

30

期数:

2020年04期

页码:

1-7

栏目:

智能、算法、系统工程

出版日期:

2020-04-10

文章信息/Info

Title:

A Face Detection Method with Lightweight and Multi-scale Feature

文章编号:

1673-629X(2020)04-0001-07

作者:

朱 鹏¹ ; 陈 虎^1* ; 李 科² ; 程宾洋³

1. 四川大学 计算机学院,四川 成都 610065; 2. 四川大学 视觉合成图形图像技术国防重点学科实验室,四川 成都 610065; 3. 四川川大智胜软件股份有限公司,四川 成都 610045

Author(s):

ZHU Peng¹ ; CHEN Hu^1* ; LI Ke² ; CHENG Bin-yang³

1. School of Computer Science,Sichuan University,Chengdu 610065,China; 2. National Key Laboratory of Fundamental Science on Synthetic Vision,Sichuan University,Chengdu 610065,China; 3. Wisesoft Co. Ltd. ,Chengdu 610045,China

关键词:

人脸检测; 轻量级; 多尺度特征; 特征融合; 全卷积神经网络; 实时检测

Keywords:

face detection; lightweight; multi-scale feature; feature fusion; full convolutional neural network; real-time detection

分类号:

TP391

DOI:

10. 3969 / j. issn. 1673-629X. 2020. 04. 001

摘要:

目前,各种 start-of-the-art 的人脸检测算法被提出,但是在不断提高精度的同时,却忽略了检测的实时性和可应用性。 针对这一问题,提出了一种轻量级、即时性和 Single-Stage 的高精度人脸检测算法。 通过将主干网络的不同尺度的特征输出到对应的检测模块中进行检测,实现不同大小人脸的检测,提高 算法精度;主干网络运用深度可分离卷积将传统的3X3 卷积核分为一个深度卷积核和 1X1 卷积核,减少计算量;检测模块中使用特征融合获得更多的上下文信息和更大的感受野,并包含目标分类和框体回归操作;采用全卷积神经网络,减少内存量,并使得网络可以输入不同尺寸的图像。 实验结果表明,算法在 FDDB 数据集上有着 93.52% 的准确率,同时在不同尺度、姿态、装扮和光照等环境下具有较好的鲁棒性,并且能够达到实时检测。

Abstract:

At present, various start-of-the-art face detection algorithms have been proposed, but while continuously improving the accuracy,the real-time and applicability of detection have been neglected. Aiming at this problem,we propose a lightweight,real-time and Single- Stage high - precision face detection algorithm. We input different scale features of the backbone network into different detection modules to realize the detection of different size faces and improve the detection accuracy of the algorithm. The backbone network uses the deep separable convolution to divide the traditional 3X3 convolution kernel into a deep convolution kernel and a 1X1 convolution kernel,reducing the amount of computation. The detection module uses feature fusion for more contextual information and greater receptive fields, including target classification and box regression operations. In addition, we use a full convolutional neural network to save memory and enable the network to input images of different sizes. The experiment shows that the proposed algorithm has 93.52% accuracy in FDDB dataset and better robustness in different scales,poses,dressing and illumination,which can achieve real-time detection.

常用功能

更新日期/Last Update: 2020-04-10