期刊名称：Journal of Theoretical and Applied Information Technology
印刷版ISSN：1992-8645
电子版ISSN：1817-3195
出版年度：2021
卷号：99
期号：13
语种：English
出版社：Journal of Theoretical and Applied
摘要：In this article we consider end-to-end full page Handwritten Text Recognition for offline Kazakh text images written in Cyrillic alphabet using Fully connected CNN and bidirectional LSTM. The model performs training of text segmentation and recognition jointly using a new Kazakh text images dataset, named Kazakh Handwritten Dataset (KHD). The novel method, which we introduce, uses three steps: Start, Follow and Read (SFR). The proposed model makes use of Region Proposal Network in order to find the starting coordinates of lines in the page. For the case when lines are not straight, we introduce a method that pursues text lines until the end of it and prepare it for the last recognition step. The SFR model works for Russian language as well since Russian alphabet is a subset of Kazakh alphabet. The experimental analysis shows that on average the model provides 0.11 Character Error Rate.
关键词：Computer Vision;HTR;CNN;Bidirectional LSTM;Kazakh Handwritten;Document Processing;Text Line Follower;Text line cutting
