基于自动提取特征的手写体数字识别  

Handwriting Digital Recognition Based on Extraction Feature Automatically

作　　者：张全 刘玉良[1] 刘国华[2] Zhang Quan;Liu Yuliang;Liu Guohua(Advanced Structural Integrity International Joint Research Centre,College of Electronic Information and Automation,Tianjin University of Science&Technology,Tianjifi 300222,China;Tianjin Key Laboratory of Optoelectronic Sensor and Sensing Netuvrk Technoloy,College of Electronic Information and Optical Engineering,Nankai University,Tianjin 300071,China)

机构地区：[1]天津科技大学电子信息与自动化学院先进结构完整性国际联合研究中心,天津300222 [2]南开大学电子信息与光学工程学院,天津市光电传感器与传感网络技术重点实验室,天津300071

出　　处：《南开大学学报（自然科学版）》2022年第1期47-51,共5页Acta Scientiarum Naturalium Universitatis Nankaiensis

基　　金：国家自然科学基金(61771261)。

摘　　要：应用基于弃权、扩展数据、最大池化技术的卷积神经网络对MNIST手写数字数据集进行识别.对深度学习模型的代价函数,训练方法进行研究,优化了传统神经网络,达到了有效自动提取特征的目的.相较于人工特征提取和传统逻辑回归分析提高了识别率,减少了错误率.实验结果表明得到了99.1%的平均识别率,减少了76%的错误率.An automatic extraction feature method is proposed which is convolution neural network(CNN)algorithm that combined with techniques that are dropout,expanding the data and maximum pooling to recognizing the hand written numerals from MNIST data set.The cost function and training method of deep learning model are studied,and the traditional neural network is optimized to achieve the purpose of effectively extracting features automatically.Compared with the results of artificial feature extraction method and traditional logistic regression analysis algorithm,the performance of the new method was improved and the error rate was reduced.Experiment results shows that the accuracy has been found to be 99.1%,the error rate has been reduced by 76%.

关 键 词：自动特征提取 数据扩展 手写体识别 深度学习 

分 类 号：TP212.3[自动化与计算机技术—检测技术与自动化装置]