A handwriting text recognition system based on neural network, comprising:
- a stroke input processor for receiving tracks of online handwriting texts, where the handwriting texts exist corresponding to character strings;
- the tracks being divided into several strokes;
- in a training stage and a prediction stage of a neural network, the stroke input processor pre-processing the strokes;
- in training stage, these strokes being corresponding to a character strings which is a known character string;
- each character string being formed by at least one character;
- a handwriting track and corresponding character string being formed as a character string sample;
- a string database connected to the stroke input processor for storing a large amount of the tracks;
- in training stage, the string database serving for storing the tracks and corresponding character strings which are formed as character string samples;
- a word recognition neural network including an encoder, a decoder, an input interface in the encoder, and an output interface in the decoder;
- the input interface being connected to the string database for receiving the strokes of the tracks;
- the encoder serving for finding a context matrix from the strokes and an initial state used by the decoder;
- the decoder receiving the context matrix and the initial state;
- furthermore the output interface serving for outputting parts of the predicted character strings and confidences of these parts, in that, in the output of the decoder, the predicted character strings being outputted part by part, that is to say, in each operation of the decoder, only a part of the predicted character strings and its confidence are outputted, and then this part and the confidence are transferred back to the input of the decoder to generate a further part of the predicted character strings and confidence thereof;
- this further part includes the former part of the predicted character strings inputted to the decoder;
- these processes are performed again until a set condition is achieved;
- wherein in the training stage, each stroke of the track is inputted to the word recognition neural network from the input interface, and the character strings corresponding to the track are used as predicted output of the word recognition neural network;
- and an after-processor being connected to the string database and the output interface of the text recognition neural network;
- the character strings in the candidate unit are normalized in the after processor to be utilized as outputs of the word recognition neural network.