基于Tensorflow深度学习框架的手写体数字识别模型优化及应用  

Optimization and Application of Handwritten Number Recognition Model Based on Tensorflow Deep Learning Framework

作　　者：涂朴 黄晨 TU Pu;HUANG Chen(School of Intelligent Manufacturing,Sichuan University of Arts and Science,Dazhou 635000 China;Dazhou Public Security Bureau,Dazhou 635000 China)

机构地区：[1]四川文理学院智能制造学院,四川达州635000 [2]四川省达州市公安局,四川达州635000

出　　处：《自动化技术与应用》2020年第12期110-114,共5页Techniques of Automation and Applications

基　　金：四川省教育厅项目(编号18ZB0508);四川文理学院校级项目(编号2018ZNZZ004Y)。

摘　　要：优化后的基于TensorFlow的全连接神经网络模型,在手写体数字数据集上的识别正确率从91.2%提升到98.3%。增加隐藏层和加入RELU激活函数,对提升识别率效果显著,而指数衰减的学习率、滑动平均模型和正则化损失则对识别正确率无明显影响。将改进后的模型,移植到ROS操作系统中,调用已训练好的模型及参数,并将识别功能封装成ROS节点,最后运用消息机制对摄像头采集到的图像消息完成识别。In this paper,the full-connected neural network model based on TensorFlow is optimized,and the recognition accuracy on handwritten digital data sets is improved from 91.2%to 98.3%.By adding hidden layer and RELU activation function,the recognition rate is improved significantly,while the learning rate of exponential attenuation,sliding average model and regularization loss have no significant impact on the recognition accuracy.The improved model is transplanted to the ROS operating system,the trained model and parameters are invoked,and the recognition function is encapsulated into ROS nodes.Finally,the image message collected by the camera is recognized by message mechanism.

关 键 词：深度学习 神经网络 TensorFlow 手写体数字识别 ROS 

分 类 号：TP391.43[自动化与计算机技术—计算机应用技术]