摘要
传统人脸检测算法在复杂环境背景下一直存在着检测准确率及效率低等问题。近年来,得益于人脸数据集的增长以及计算机硬件的极速发展,使用深度神经网络的人脸检测算法在准确度方面已有很大提升,但使用的模型结构越来越复杂,检测速度也相对变慢。本文提出一种改进的多任务卷积神经网络(Multi-task convolutional neural networks,MTCNN)算法。在制造数据集时更改IOU阈值参数,来获取更多、更精确的人脸样本;对与置信度损失有关的交叉熵损失函数和与偏移量损失有关的均方差函数求均值,使得整个网络收敛得更加平稳。经在AFW、PASCAL以及FDDB数据集上实验,与传统算法相比,该算法在保证实时性的同时提升了检测准确率,可应用于追求更高准确率的人脸检测系统。
Traditional face detection algorithms have always had problems such as low detection accuracy and low efficiency in the context of complex environments.In recent years,it has benefited from the growth of face data sets and the rapid development of computer hardware.Face detection algorithms using deep neural networks are in accuracy.This aspect has been greatly improved,but the structure of the model used is becoming more and more complex.The detection speed is relatively slow.Therefore,it is very important to design a detection model that takes into account both accuracy and real-time performance.This paper proposes an improved multi-task convolutional neural networks(MTCNN)algorithm.The first in the manufacturing data sets changes the IOU threshold parameter to obtain a more accurate face samples.Secondly,the cross emotion loss function related to the confidence loss and the average of the mean square error function related to the offset loss make the convergence of the entire network more stable.After experiments on the AFW,PASCAL and FDDB data sets,compared with traditional algorithms,this algorithm improves the detection accuracy while ensuring real-time performance,which could be applied to face detection systems pursuing higher accuracy.
作者
周航
蔡茂国
吴涛
沈冲冲
ZHOU Hang;CAI Maoguo;WU Tao;SHEN Chongchong(College of Electronics and Information Engineering,Shenzhen University,Shenzhen Guangdong 518060,China)
出处
《智能计算机与应用》
2021年第3期172-176,共5页
Intelligent Computer and Applications
