摘要
该文首次将ResNet网络的思想对复杂教室环境下的人物进行多类别分类设计,改进了网络结构,有效解决了传统基于像素特征的方法分类效果不理想的问题。实验中通过卷积提取特征、不同感受野、保留像素间联系、多层卷积级联提取深层次特征等方法,在网络训练过程中进行参数调整,优化算法和网络参数来解决困难样本的识别,将多类别的分类准确率从83.5%提升到99.2%,并实现了多目标检测的11类样本的判定。最终选取ResNet1816来进行高速有效的多类别识别。
The multiclass classification for the students in the complex classroom environment is designed with the idea of ResNet network,which improves the network structure,and effectively solves the problem that the classification effect is not ideal in the traditional method based on pixel feature. In the experiments,the convolution is used for the feature extraction,which adopts the different receptive fields,preserves the connections between pixels,uses the multilayer convolution cascade to extract the deep-seated features and other methods. The parameter adjustment is conducted in the network training process,and the algorithm and network parameters are optimized to identify the difficult samples,which increase the accuracy of multiclass classification from 83.5% to 99.2%,and realize the determination of 11 kinds of samples of multi-target detection. In this paper,Resnet1816 is finally selected for the high-speed and effective multiclass recognition.
作者
倪照风
马原东
崔潇
郦烜杰
杨秀璋
罗子江
NI Zhaofeng;MA Yuandong;CUI Xiao;LI Xuanjie;YANG Xiuzhang;LUO Zijiang(School of Information,Guizhou University of Finance and Economics,Guiyang 550025,China;Beijing Interjoy Technology Limited Company,Beijing 100089,China;School of Electronics,University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《现代电子技术》
北大核心
2020年第12期42-46,共5页
Modern Electronics Technique
基金
国家自然科学基金项目(11664005)。
关键词
姿态分类
ResNet网络
卷积网络
分类训练
参数调整
多类别识别
gesture classification
ResNet network
convolution network
classification training
parameter adjustment
multiclass identification
