摘要
为提升面向嵌入式设备的人脸性别识别和年龄段估计任务的鲁棒性和检测效率,提出了一种共享浅层卷积特征的多任务学习网络。以裁剪后的EfficientNet为卷积特征提取层,构建多任务学习卷积神经网络(CNN)模型;并将年龄判别中的序数回归问题转换为多个二分类子问题以优化损失函数;通过将系统部署在NVIDIA TX2嵌入式平台上,验证了改进后算法识别准确率提高了7.95%,运行速率可达30 fps。
In order to improve the robustness and detection efficiency of facial gender recognition and age estimation tasks for embedded devices,a multi-task learning network sharing shallow layer convolutional features is proposed.The network takes the clipped EfficientNet as convolutional feature extraction layer,and a multi-task learning convolutional neural network(CNN)model is constructed.In order to optimize the loss function,the ordinal regression problem in age discrimination is transformed into several dichotomous sub-problems.By deploying the system on NVIDIA TX2 embedded platform,the improved recognition accuracy of the algorithm is increased by 7.95%,and the running rate can reach 30 fps.
作者
罗洪伟
刘波
姚虎
王建博
袁华清
刘桂华
LUO Hongwei;LIU Bo;YAO Hu;WANG Jianbo;YUAN Huaqing;LIU Guihua(Shenzhen Langchi Xinchuang Technology Co Ltd,Shenzhen 518000,China;School of Electrical Automation and Information Engineering,Tianjin University,Tianjin 300072,China;School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China)
出处
《传感器与微系统》
CSCD
北大核心
2023年第11期114-118,共5页
Transducer and Microsystem Technologies
关键词
深度学习
卷积神经网络
性别识别
年龄估计
表情识别
deep learning
convolutional neural network(CNN)
gender recognition
age estimation
expression recognition
