End-to-end Text Recognition Model Based on Handwritten Chinese Characters

Abstract:
Handwritten text recognition is widely used in document transcription and document digitization. The existing text recognition models mainly use the detection model to detect the text and then use the recognition model to recognize the text. This paper designs an end-to-end model that integrates the detection module and the recognition module for text images, which can save a lot of network model parameters and shorten the calculation time. The main feature extraction, detection and recognition of the network model, in which the detection module and the recognition module share a convolutional neural network, to achieve the purpose of parameter sharing and reducing the amount of calculation. The model can locate and recognize text information through one forward propagation. Compared with traditional text recognition models, the weights of convolution features do not need to be calculated separately during training and inference, thus saving the time of forward propagation. At the same time, data enhancement methods are introduced for handwritten Chinese characters to increase the generalization ability of the model, the Long Short-Term Memory network is integrated into the text recognition part. According to the given context, the accuracy of handwritten Chinese character recognition is effectively improved.