基于改进MTCNN的人脸检测算法  

Face detection algorithm based on improved MTCNN

作　　者：蓝雯飞[1] 张盛兰 朱容波[1] 熊文娟 LAN Wenfei;ZHANG Shenglan;ZHU Rongbo;XIONG Wenjuan(College of Computer Science,South-Central University for Nationalities,Wuhan 430074,China)

机构地区：[1]中南民族大学计算机科学学院,武汉430074

出　　处：《中南民族大学学报（自然科学版）》2020年第6期637-641,共5页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：国家自然科学基金资助项目(61772562)。

摘　　要：针对经典的人脸检测卷积神经网络模型(CNN)对人脸检测准确率不高的问题,设计一种改进的多任务卷积神经网络(MTCNN).通过对原始MTCNN算法进行迁移学习,微调模型参数,提出人脸误检判别公式.在LFW人脸数据集上进行实验,先调整MTCNN关键参数的值,找出最合适的人脸置信度阈值,再使用人脸误检判别公式.实验结果表明:改进后的MTCNN算法较改进之前在检测准确率上有了较大提升,而且提出的改进策略使自然环境中人脸检测的速度也有了提高.Aiming at the problem that the classical face detection convolutional neural network model(CNN)has low accuracy of face detection,an improved multi-task convolutional neural network(MTCNN)is designed.Through the transfer learning of the original MTCNN algorithm,fine-tuning the model parameters,a facial misdetection discrimination formula is proposed.Experiment on the LFW face data set,first adjust the value of MTCNN key parameters,find the most suitable face confidence threshold,and then use the face misdetection discrimination formula.Experimental results show that the improved MTCNN algorithm has greatly promoted the detection accuracy rate compared to before,and the proposed improvement strategy has also improved the speed of face detection in natural environment.

关 键 词：人脸检测 卷积神经网络 MTCNN算法 迁移学习 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]