摘要：

现有的性别与年龄检测系统一般都需要大量的运算,难以有效地集成到嵌入式系统中,且目前公开的数据集中没有大型东亚人脸数据库,使用公开的西方人数据库训练出的性别与年龄模型在检测东亚人脸数据集时效果并不理想。为此,提出一种改进的DeepID网络模型。在互 联网上收集并整理3万多张不同性别与年龄段的东亚人脸数据集,并将该数据集用于训练新的神经网络,通过训练改进的第一代DeepID网络,可在不降低网络分类精度的前提下,提高嵌入式系统性别与年龄分类速度。实验结果表明,该模型可以有效地运行在嵌入式系统,且在检 测含东亚人脸的数据集时检测精度明显提高。

关键词: 深度学习, 卷积神经网络, 模式识别, 性别识别, 年龄识别, 嵌入式系统

Abstract:

Existing gender and age detection system generally requires a large amount of calculations and is difficult to integrate into embedded systems effectively.There is no large-scale East Asian human face database in the currently open data set,and gender and age trained using an open western database are used.The model is not ideal for detecting East Asian face data sets.Therefore,an improved DeepID network model is proposed.More than 30 000 East Asian face data sets of different genders and age groups are collected and organized on the Internet.The data set is used to train new neural networks.The first generation DeepID network is improved through training,improve the gender and age classification speed of the embedded devices without reducing the accaracy of network classification.Experimental results show that the model can be effectively run on the embedded end,and the detection accuracy is significantly improved when detecting datasets with East Asian faces.

Key words: deep learning, Convolutiona Neural Network(CNN), pattern recognition, gender recognition, age recognition, embedded system

中图分类号: 

• TP18