Handwritten Uyghur Text Recognition Method Based on Printed Text Images Supervision
A handwritten Uyghur character image dataset is scarce and handwritten text is difficult to segment and recognize.A handwritten Uyghur character recognition model based on Printed text image Supervision(PSnet)was proposed.This model regards text and printed text images as labels,and during the training phase they are put into Convolutional Neural Network(CNN)to extract features in a parallel way.Then the features are put into recognition task and match task at the same time.During the testing phase,the features are put into BiLSTM encoder to acquire per frame prediction.Finally the Connectionist Temporal Classification(CTC)decoder is used to obtain the final pre-diction result.The experiment results show that the method is able to generate plenty of effective handwritten text ima-ges and PSnet gets accuracy of 0.44%and 5.1%CER on seen and unseen fonts test datasets,and gets 19%CER which reduces by 5.03%on artificial handwritten test datasets compared to other base models.The results implicate the method can relieve the data rarity and utilize the printed text images features to assist handwritten Uyghur recogni-tion.
万方数据
维普
