Laser & Optoelectronics Progress, Volume. 58, Issue 12, 1210007(2021)
Offline Handwritten Text Recognition Based on CTC-Attention
Aiming at the problems of casual writing of the offline handwritten text, difficulty in character segmentation, and the dependence of recognition accuracy on a dictionary, an offline handwritten text recognition algorithm based on connectionist temporal classification (CTC)-attention is proposed. The convolutional neural network and bidirectional long short-term memory are used to encode the image features. Multitask learning framework based on CTC and Attention-based models is used to decode feature sequences. In the training process, the CTC model and the attention mechanism model are used to train at the same time, which effectively solves the problem of ignoring the overall information when CTC predicts local information, and the problem of unconstrained decoding of the attention mechanism.Experiments on IAM dataset, i.e., the classical handwritten English word dataset, showed that the character accuracy rate of the proposed method is 93.4%, and the word accuracy rate is 81.8%, proving the proposed method’s feasibility.
Get Citation
Copy Citation Text
Yangyang Ma, Bing Xiao. Offline Handwritten Text Recognition Based on CTC-Attention[J]. Laser & Optoelectronics Progress, 2021, 58(12): 1210007
Category: Image Processing
Received: Aug. 24, 2020
Accepted: Oct. 14, 2020
Published Online: Jun. 18, 2021
The Author Email: Xiao Bing (16392603@qq.com)