Method, device and server for recognizing characters of claim document, and storage medium

A server for recognizing characters of a claim document, comprising a memory and a processor, wherein the memory stores a program of recognizing the characters of the claim document, and the program of recognizing the characters of the claim document is executed by the processor to realize the following steps:

• after the server receives a claim document image to be subjected to a character recognition, performing a region segmentation according to a frame line layout of a claim document frame format, thus obtaining one or multiple segmented regions;
• calling a predetermined analysis model to analyze all the obtained segmented regions, and performing the character recognition on all the analyzed segmented regions respectively by using a predetermined recognition rule to recognize the characters in all the segmented regions;
• wherein the step of calling the predetermined analysis model to analyze all the obtained segmented regions comprises:
• calling the predetermined analysis model to analyze all the obtained segmented regions so as to obtain first segmented regions which are recognized by an optical character recognition engine and second segmented regions which are not recognized by the optical character recognition engine;
• wherein the step of performing the character recognition on all the analyzed segmented regions by using the predetermined recognition rule further comprises:
• performing the character recognition on all the first segmented regions by using the predetermined optical character recognition engine to recognize characters in all the first segmented regions, and calling a predetermined recognition model to perform the character recognition on all the second segmented regions to recognize characters in all the second segmented regions;
• wherein the predetermined recognition model is an LSTM (Long Short-Term Memory) model, and the training process of the predetermined recognition model is as follows:
• obtaining a preset number of segmented region samples, and labeling all the segmented region samples by characters included in these segmented region samples;
• classifying the preset number of segmented region samples into a first data set and a second data set according to a preset ratio, taking the first data set as a training set, and taking the second data set as a test set;
• sending the first data set into an LSTM network for a model training, performing the character recognition on the segmented region samples in the second data set by using the trained model every preset time, and comparing recognized characters with labels of the segmented region samples so as to calculate errors between the recognized characters and the labels;
• if the errors of the characters recognized by the trained model are diverged, adjusting preset training parameters for retraining till the errors of the characters recognized by the trained model are converged;
• if the errors of the characters recognized by the trained model are converged, ending the model training, and taking the generated model as the trained predetermined recognition model.

A non-transitory computer readable storage medium, wherein the computer readable storage medium stores a program of recognizing the characters of the claim document, and the program of recognizing the characters of the claim document is executed by at least one processor to realize the following steps:

• after a server receives a claim document image to be subjected to a character recognition, performing a region segmentation according to a frame line layout of a claim document frame format, thus obtaining one or multiple segmented regions;
• calling a predetermined analysis model to analyze all the obtained segmented regions, and performing the character recognition on all the analyzed segmented regions respectively by using a predetermined recognition rule to recognize the characters in all the segmented regions;
• wherein the step of calling the predetermined analysis model to analyze all the obtained segmented regions comprises:
• calling the predetermined analysis model to analyze all the obtained segmented regions so as to obtain first segmented regions which are recognized by an optical character recognition engine and second segmented regions which are not recognized by the optical character recognition engine;
• wherein the step of performing character recognition on all the analyzed segmented regions by using the predetermined recognition rule further comprises:
• performing the character recognition on all the first segmented regions by using the predetermined optical character recognition engine to recognize characters in all the first segmented regions, and calling a predetermined recognition model to perform the character recognition on all the second segmented regions to recognize characters in all the second segmented regions;
• wherein the predetermined recognition model is an LSTM (Long Short-Term Memory) model, and the training process of the predetermined recognition model is as follows:
• obtaining a preset number of segmented region samples, and labeling all the segmented region samples by characters included in these segmented region samples;
• classifying the preset number of segmented region samples into a first data set and a second data set according to a preset ratio, taking the first data set as a training set, and taking the second data set as a test set;
• sending the first data set into an LSTM network for a model training, performing the character recognition on the segmented region samples in the second data set by using the trained model every preset time, and comparing recognized characters with labels of the segmented region samples so as to calculate errors between the recognized characters and the labels;
• if the errors of the characters recognized by the trained model are diverged, adjusting preset training parameters for retraining till the errors of the characters recognized by the trained model are converged;
• if the errors of the characters recognized by the trained model are converged, ending the model training, and taking the generated model as the trained predetermined recognition model.