作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2021, Vol. 47 ›› Issue (7): 273-280. doi: 10.19678/j.issn.1000-3428.0057249

• 图形图像处理 • 上一篇    下一篇

基于轻量级卷积神经网络的人脸检测算法

朱灵灵1, 高超2, 陈福才2   

  1. 1. 郑州大学 中原网络安全研究院, 郑州 450000;
    2. 中国人民解放军战略支援部队信息工程大学, 郑州 450000
  • 收稿日期:2020-01-16 修回日期:2020-05-29 发布日期:2020-06-09
  • 作者简介:朱灵灵(1995-),女,硕士研究生,主研方向为计算机视觉、图像处理;高超,助理研究员;陈福才,研究员。
  • 基金资助:
    国家自然科学基金(61601513)。

Face Detection Algorithm Based on Lightweight Convolutional Neural Network

ZHU Lingling1, GAO Chao2, CHEN Fucai2   

  1. 1. Zhongyuan Network Security Research Institute, Zhengzhou University, Zhengzhou 450000, China;
    2. PLA Strategic Support Force Information Engineering University, Zhengzhou 450000, China
  • Received:2020-01-16 Revised:2020-05-29 Published:2020-06-09

摘要: 针对人脸检测在移动端应用时面临的移动设备计算能力及存储资源受限等问题,设计一种基于轻量级卷积神经网络的改进人脸检测算法Lightweight-SSH。基于单点无头人脸检测器(SSH)人脸检测算法,采用基于MobileNet的轻量级卷积神经网络对样本数据进行特征提取,减少模型的参数量和计算量,通过在SSH网络的检测模块中引入可变形卷积层,提升卷积神经网络对人脸形变的建模能力。在Wider Face数据集上的实验结果表明,与常用人脸检测算法相比,Lightweight-SSH算法在保证检测精度的前提下,明显降低模型复杂度,并提高了模型检测速度。

关键词: 人脸检测, 可变形卷积, MobileNet网络, 单点无头人脸检测器, Lightweight-SSH算法

Abstract: When deployed to mobile terminals,face detection applications are usually limited by the computing power and storage resources of the devices.To address the problem,an improved face detection algorithm called Lightweight-SSH using lightweight convolutional neural network is proposed. This face detection algorithm is designed based on the Single Stage Headless Face Detector(SSH) algorithm,and employs MobileNet-based lightweight convolutional neural network to extract the features of sample data,reducing the number of parameters and the amount of computation.A deformable convolution layer is introduced into the detection module of the SSH network to improve the ability of the convolutional neural network to model human face deformation.Experimental results on the Wider Face dataset show that,compared with commonly used face detection algorithms,the Lightweight-SSH algorithm significantly reduces model complexity and improves model detection speed while ensuring detection accuracy.

Key words: face detection, deformable convolution, MobileNet, Single Stage Headless Face Detector(SSH), Lightweight-SSH algorithm

中图分类号: