Cross-domain Author Adaptive Handwritten Recognition Based on CNN-ISVM
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 CNN-ISVM,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 ex-periment,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.
万方数据
维普
