面向手写汉字识别的残差深度可分离卷积算法  

Residual Depthwise Separable Convolution Algorithm for Handwritten Chinese Character Recognition

作　　者：陈鹏飞 应自炉[1] 朱健菲 商丽娟 CHEN Peng-fei;YING Zi-lu;ZHU Jian-fei;SHANG Li-juan(School of Information Engineering,Wuyi University,Jiangmen 529020,China)

机构地区：[1]五邑大学信息工程学院,广东江门529020

出　　处：《软件导刊》2018年第11期68-72,76,共6页Software Guide

基　　金：国家自然科学基金项目(61771347;61372193)

摘　　要：卷积神经网络(CNN)在脱机手写汉字识别领域取得了杰出的研究成果。采用更深层卷积神经网络可取得更高识别准确率,但是模型容量和计算复杂度也会增加,将模型部署到存储资源和计算力有限的移动设备上变得更加困难。为解决上述问题,提出一种基于深度可分离卷积的残差卷积神经网络。深度可分离卷积将标准卷积操作分离成特征提取和特征融合,逐深度卷积被用于特征提取,特征融合采用逐点卷积实现。使用深度可分离卷积改进残差网络,实现较深层的残差网络。模型使用联合的中心损失函数和softmax损失函数进行监督训练,可使模型学习具有判别性特征,提高了模型识别准确率。采用CASIA-HWDB数据集进行实验,结果表明该方法具有较低的模型容量和计算复杂度,能够达到96.50%的主流识别率。Convolutional neural networks have achieved many outstanding results in the field of offline handwritten Chinese character recognition.Deeper network models lead to higher recognition accuracy,but deeper models consume more storage and more computational resources.This makes it more difficult to deploy the model to resource-constrained computing devices.In order to solve the problem,an efficient convolutional neural network based on depthwise separable convolution is proposed.Depthwise separable convolutional network separates the standard convolution operation into two parts:feature extraction and feature fusion,depthwise convolution is used for feature extraction and feature fusion is implemented by pointwise convolution.Depthwise separable convolution is used to improve the residual network,achieve a deeper residual network.joint supervised training of center loss function and softmax loss function is used.Joint softmax loss function and center loss function can make models learning discriminative features and improve the model recognition accuracy.Experiments on the CASIA-HWDB dataset are carried out,the proposed approach has lower storage and computational complexity achieves the state-of-art recognition accuracy of 96.50%.

关 键 词：脱机手写汉字识别 残差卷积神经网络 深度可分离卷积 中心损失 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]