GRU-CNN优化模型的手势动作识别研究  

Research on Gesture Action Recognition Based on GRU-CNN Optimization Model

作　　者：刘慧婷 李建军[1] LIU Hui-ting;LI Jian-jun(College of Information Engineering,Inner Mongolia University of Science and Technology,Baotou 014010,China)

机构地区：[1]内蒙古科技大学信息工程学院,内蒙古包头014010

出　　处：《小型微型计算机系统》2021年第10期2121-2124,共4页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(62066036)资助;内蒙古自然科学基金项目。

摘　　要：近年来,动态手势识别已经成为计算机视觉领域的研究热点.由于手势动作的复杂性和多样性导致动态手势识别的识别率不是很高.随着深度学习的发展,卷积神经网络(CNN)在图像视频领域方面表现出较好的性能,但是,卷积神经网络仅仅考虑了图像序列的当前输入,丢失了上下文关系,为了解决这一问题,鉴于门控循环单元(GRU)网络具有对有长时间关联性数据较强的处理能力,本文提出了GRU-CNN融合网络模型.该模型可以提取更多的时空信息,而且具有较少的网络参数,收敛时间短,能更好的满足实时性需求.本模型在公开数据集MSRC-12上取得了良好的实验效果,同时分析了超参数(批量大小,权重和偏差学习率)对分类精度的影响.在网络模型的对比实验中,GRU-CNN的性能大大优于CNN-GRU模型.In recent years,dynamic gesture recognition has become a research hotspot in the field of computer vision.Due to the complexity and diversity of gesture actions,the recognition rate of dynamic gesture recognition is not very high.With the development of deep learning,convolutional neural networks(CNN) have shown better performance in the field of image and video.However,convolutional neural networks only consider the current input of the image sequence and lose the context.In order to solve this problem,in viewof The gated recurrent unit(GRU) network has strong processing capabilities for long-term correlation data.This paper proposes a GRU-CNN fusion network model.The model can extract more spatio-temporal information,has fewer network parameters,has a short convergence time,and can better meet real-time requirements.The influence of hyperparameters(batch size,weight and deviation learning rate) on classification accuracy is analyzed.This model has achieved good experimental results on the public data set MSRC-12.In the comparison experiment of network models,the performance of GRU-CNN is much better than that of CNN-GRU model.

关 键 词：图像处理 深度学习 动态手势识别 卷积神经网络 门控循环单元 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]