基于CNN-ISVM的跨领域书写人自适应手写识别  

Cross-domain Author Adaptive Handwritten RecognitionBased on CNN-ISVM

作　　者：张墨逸[1] 叶洪昶 袁小芳 陈海燕[1] ZHANG Mo-yi;YE Hong-chang;YUAN Xiao-fang;CHEN Hai-yan(School of Computer and Communication,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]兰州理工大学计算机与通信学院,甘肃兰州730050

出　　处：《计算机技术与发展》2024年第12期187-193,共7页Computer Technology and Development

基　　金：国家自然科学基金项目(62161019)。

摘　　要：用户书写风格、字形、笔迹、书写方式等方面都会存在差异,使手写识别应用具有特异性、小样本和多样式的特点,普适性的模型很难满足,必须在手写过程中对特定用户的书写进行自适应学习,使模型更好地服务于人们的个性化需求。针对此问题,研究者提出基于CNN-ISVM的跨领域书写人自适应手写识别方法,灵活地调整普适化模型,进行个性化的手写识别。在构造通用模型时,利用CNN作为特征提取器,对图像进行特征学习和提取,将提取的特征输入到SVM中进行分类。自适应手写识别时,引入基于错分样本触发的ISVM增量学习方法,使用增量样本和保存好的支持向量集对模型在线学习和更新。在实验中,当源域由静态手写图片组成,目标域为2组空写数据集时,每类别增量5张样本,识别率分别达到92.8%、90.42%。该方法简单易行,可以在目标域每类样本数据只有1张增量样本的情况下进行跨领域书写人自适应学习,与其它方法相比,识别率有较明显的提升。Different people have differences in writing style,font shape,script,writing method and etc.,so the hand writing recognition application has the character of specificity,small sample size,and multi-style.The universal model cannot meet the writing requirements of a specific usage,so it is necessary to self-adapt the writing of a specific user in the handwriting process,to make the model better serve people’s personalized needs.In order to solve the above problem,we propose a cross-domain handwritten recognition method based on CNNISVM,which flexibly adjusts the universal model and carries out individuation handwriting recognition.In the construction of the generalized model,CNN is used as a feature extractor to learn and extract features from handwritten images,then CNN extracted features are input into SVM classifier for classification.In the author adaptive recognition,the wrongly predicted sample is introduced to trigger the incremental processing,which can use incremental set and save support vector set for online learning and updating the model.In experiment,when the source domain is composed of static handwritten images,and the target domains are two in-air handwritten datasets,with incremental images 5 per category,the recognition rates are 92.8%and 90.42%respectively.The proposed method is simple and easy to implement,which can implement cross-domain author adaptive learning when there is only one incremental sample data in the target domain for each class.Compared with other methods,the recognition accuracy of the proposed method is significantly improved.

关 键 词：手写识别 书写人自适应 领域自适应 增量学习 卷积神经网络 支持向量机 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]