基于改进的RetinaFace快速单一人脸检测算法研究
Research on Improved RetinaFace Fast Single Face Detection Algorithm
胡胜林 1丁健 1张火强 1汪慧1
作者信息
- 1. 合肥大学 先进制造工程学院,安徽 合肥 230000
- 折叠
摘要
在一些特殊的场景中需要进行快速单一的人脸检测,例如身份认证、人脸追踪、疲劳驾驶检测等.针对此类情况,通过研究分析轻量级网络,提出一种基于RetinaFace的检测方法.首先,选用高效的轻量级网络MobileNetV3 搭建主干网络;其次,在主干网络与FPN层之间融入ECANet网络架构加强特征提取能力;最后,对FPN层进行调整以提高检测速度.实验结果表明,改进后的算法模型在WiderFace简单和中等程度子集上的平均精度分别提高了 3.6%和5.7%,检测速度提高了21.6%和13.1%.
Abstract
In certain specialized scenarios,there is a need for rapid and singular facial detection,such as i-dentity verification,facial tracking,and fatigue detection during driving.To address these problems,this paper proposes a detection method based on RetinaFace through the study and analysis of lightweight networks.Initial-ly,an efficient lightweight network,MobileNetV3,is selected to construct the backbone network.Subsequently,an ECANet is introduced between the backbone network and the FPN(Feature Pyramid Network)layer to en-hance feature extraction capabilities.Finally,adjustments are made to the FPN layer to improve detection speed.Experimental results indicate that the improved algorithm model shows an average precision increase of 3.6%and 5.7%on the simple and moderately challenging subsets of the WiderFace dataset respectively.Addi-tionally,the detection speed is enhanced by 21.6%and 13.1%.These results demonstrate that the proposed RetinaFace-based detection method,leveraging MobileNetV3 as the backbone network and incorporating ECANet for enhanced feature extraction,achieves notable improvements in both precision and speed,particularly in sce-narios like identity verification,facial tracking,and fatigue detection during driving.
关键词
人脸检测/轻量级网络/RetinaFace/MobileNetV3/ECANet/FPNKey words
face detection/lightweight network/RetinaFace/MobileNetV3/ECANet/FPN引用本文复制引用
出版年
2024
NETL
NSTL
万方数据