AN END-TO-END CRNN AND CTC APPROACH FOR OFFLINE HANDWRITTEN CHINESE TEXT RECOGNITION
TL;DR
Experiments demonstrate that the proposed approach achieves a Character Error Rate (CER) of 16,85%, corresponding to an accuracy of 83,15%, confirming the effectiveness of the CNN–BiLSTM–CTC framework for offline handwritten Chinese text recognition.
Abstract
Offline handwritten Chinese text recognition remains a challenging problem due to the large number of character classes, complex character structures, and high variability in writing styles. This paper proposes an end-to-end offline handwritten Chinese text recognition system based on a CNN–BiLSTM–CTC architecture. A convolutional neural network (CNN) is used to extract spatial features from handwritten text images, while a bidirectional long short-term memory (BiLSTM) network captures contextual dependencies between characters in both forward and backward directions. The Connectionist Temporal Classification (CTC) framework is applied to enable segmentation-free training and decoding of character sequences directly from input images. Experiments conducted on the CASIA-HWDB 2.2 offline handwritten Chinese text dataset demonstrate that the proposed approach achieves a Character Error Rate (CER) of 16,85%, corresponding to an accuracy of 83,15%, confirming the effectiveness of the CNN–BiLSTM–CTC framework for offline handwritten Chinese text recognition.