摘要
胶囊网络识别重叠手写数字时没考虑输出字符的顺序。为了用胶囊网络实现有序重叠手写数字识别,对原有的胶囊网络结构做出改进,构造了含有多个数字胶囊层的识别模型,每个数字胶囊层识别一个分类标签,然后通过动态路由机制算法更新数字胶囊层的参数,实现有序重叠手写数字识别。在构建的重叠手写数字数据集上进行测试并与卷积神经网络算法进行了比较。改进胶囊网络识别有序重叠手写数字的准确率达到87. 62%,相比于卷积神经网络准确率有了显著提升,证明了改进方法的有效性。
The capsule network cannot recognize overlapped handwritten numerals. In order to realize the orderly overlapped handwritten numeral recognition, an improved original capsule network is presented and a recognition model containing multiple digital capsule layers is constructed. Each of the capsule layers identifies a classification label. And thenupdate the parameters of the digital capsule layers through a dynamic routing mechanism algorithm to complete the orderly handwritten numeral recognition. The test recognition accuracy is over 87. 62% on the dataset of the over? lapped handwritten numerals compared with the convolutional neural network , the recognition accuracy is significantly improved. The effectiveness of the method is proved.
作者
朱娟
陈晓
ZHU Juan;CHEN Xiao(School of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China;Jiangsu Collaborative Innovation Center of Atmospheric Environment and Equipment Technology,Nanjing Information Science and Technology University, Nanjing 210044, China)
出处
《激光杂志》
北大核心
2019年第7期43-46,共4页
Laser Journal
基金
江苏省自然科学基金(No.BK20161536)
江苏省“333高层次人才培养工程”
江苏高校优势学科建设工程资助项目资助
关键词
胶囊网络
重叠手写数字
手写数字识别
图像识别
卷积神经网络
capsule network
overlapped handwritten numeral
handwritten numeral recognition
image recogni tion
convolutional neural networks