摘要
基于传统的浅层学习网络由于过度依赖于人工选择手势特征,因此不能实时适应复杂多变的自然场景。在卷积神经网络架构的基础上,提出了一种改进的多尺度深度网络手势识别模型,该模型能够利用卷积层自动学习手势特征,进而除去人工提取特征的弊端。该方法引入自适应多尺度特性来实现同一卷积层不同尺寸卷积核生成不同尺度特征,并通过级联浅层和深层的特征来达到不同抽象程度的特征图融合。同时,为了增强模型的泛化能力,提出了基于正则化约束的损失函数。实验结果表明,所提网络模型的识别精度高于普通单尺度卷积神经网络结构的识别精度,弥补了提取特征不够精细、全面及稳定性欠佳等缺点,同时网络训练所需的时间并没有大幅度增加。
Since the traditional shallow learning networks rely too much on manual selection of gesture features,they cannot adapt to complex and varied natural scenes in real time.Based on the convolutional neural network architecture,this paper proposes an improved multi-scale deep network gesture recognition model,which makes it possible to overcome the drawbacks of ma-nual extraction features by using the convolutional layer to automatically learn gesture features.In this method,the adaptive multi-scale features are introduced to realize that convolution kernels with different sizes at the same convolutional layer to gene-rate different scale features,and achieves feature map fusion with different levels by cascading shallow and deep features.In addition,in order to enhance the generalization ability of the model,this paper proposes a loss function based on regularization constraints.The experimental results show that the recognition accuracy of the proposed network model is higher than that of the ordinary single-scale convolutional neural network,and the shortcomings of imprecise and incomprehensive extraction as well as poor stability are overcome,and the time required for network training is not greatly increased.
作者
景雨
祁瑞华
刘建鑫
刘朝霞
JING Yu;QI Rui-hua;LIU Jian-xin;LIU Zhao-xia(School of Software,Dalian University of Foreign Languages,Dalian,Liaoning 116044,China)
出处
《计算机科学》
CSCD
北大核心
2020年第6期180-183,共4页
Computer Science
基金
国家自然科学基金(61501082)
大连外国语大学科研基金项目(2018XJYB27)
辽宁省教育厅科学研究一般项目(2019JYT01,2019JYT07)
辽宁省自然科学基金项目(20180550018)
辽宁省博士科研启动基金项目(2019BS061)。
关键词
手势识别
深度学习
多尺度
卷积特征
正则化
损失函数
Gesture recognition
Deep learning
Multi-scale
Convolution feature
Regularization
Loss function
