SqueezeNet和动态网络手术的脱机手写汉字识别  

Offline Handwritten Chinese Character Recognition of SqueezeNet and Dynamic Network Surgery

作　　者：周於川 谭钦红[1] 奚川龙 ZHOU Yu-chuan;TAN Qin-hong;XI Chuan-long(College of Communication and Information Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065

出　　处：《小型微型计算机系统》2021年第3期556-560,共5页Journal of Chinese Computer Systems

基　　金：信号与信息处理重庆市市级重点实验室建设项目(CSTC2009CA2003)资助。

摘　　要：传统和基于CNN的脱机手写汉字识别模型多数是为了追求更高准确率,未重视模型体积大小,模型中存在大量冗余参数,模型训练周期长并且很难在资源有限的平台上运行.针对这些问题,本文提出改进的SqueezeNet模型,保留了用小卷积核替代大卷积核的策略,采用层间的特征融合算法和L2范数约束的Softmax分类函数;然后再对参数裁剪进一步压缩,避免裁剪掉重要参数而损失过多准确率,采用动态网络手术算法来保证将误删重要参数重新拼接.并将改进后的模型与其它模型在测试集ICDAR-2013下进行对比,本文模型参数变少、训练速度快并且可移植性强,模型大小为3.2MB,在测试集ICDAR-2013中其准确率达到96.03%,对输入图预处理后再训练所得模型准确率达到96.32%.Traditional and CNN based off-line handwritten Chinese character recognition models are mostly in pursuit of higher accuracy,but ignores the network model parameters,which causes problems such as too many parameters,long training cycles,and inability to run on platforms with limited hardware resources;In order to solve these problems,a strategy of improving the SqueezeNet model and retaining small convolution kernels instead of large one is proposed,In addition,the feature fusion algorithm between layers and the Softmax function with L2-norm constraints are used.To further trim parameters to avoid cutting important parameters,a dynamic network surgery algorithm is used to ensure that important parameters are deleted by mistake are rejoined.By comparing the improved model with other models under the test set ICDAR-2013,the model parameters become less,the training becomes faster,and the portability is strong.The model size is 3.2 MB.In the test set ICDAR-2013,the accuracy rate reached 96.03%,and when the input image was preprocessed,it will reach to 96.32%.

关 键 词：SqueezeNet 动态手术网络 脱机手写汉字识别 深度学习 特征融合 L2-Softmax 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]