Chinese Picture Text Extraction And Error Correction Based On Deep Learning

As a task to automatically check and correct Chinese sentences,Chinese text error correction mainly aims at enhancing the standardization of language and characters while reducing the cost of manual verification.As a result,a key research topic in the field of natural language processing lies in how to quickly find out the errors in the text and correct them accurately.In terms of the text error correction task at present,most researchers directly make the exploration by focusing on the document text,who have ignored that the transformation of characters to text is a time-consuming and error prone process in practical application scenarios.In order to solve the problem of image and text content review in new media platforms,this paper proposes a deep learning model Rep OCR is proposed for optical character recognition and a Chinese error detection and correction model based on the combination of ERNIE and Bi LSTM-CRF sequence annotation,integrating Chinese text automatic proofreading from image text extraction to text error detection and correction.Chinese text error correction extends to text in pictures,not limited to text documents.The following is the main research contents of this thesis:(1)The method of extracting image text from optical character recognition is studied.A Rep OCR model is proposed,in which a re-parameterized backbone network is included,so that the multi-channel structure of the training network can be converted into a one-way structure of the reasoning network,and the network can have an efficient reasoning rate.In addition to the backbone network,Group FPN neck structure is added behind the backbone network for feature fusion,so as to realize the better integration of the feature information of small-size targets with that of large-size targets.What’s more,the clustering method is used to generate the size and aspect ratio of the prediction box of the text,so the recognition ability of the model for long text can be enhanced.(2)The Chinese text error detection methods are studied.The current mainstream sequence tagging deep learning model for text error correction is referred to,an error detection framework based on the combination of ERNIE and Bi LSTM-CRF sequence annotation is proposed,ERNIE is used for text vectorization,Bi LSTM bidirectional structure extracts contextual information and concatenates it to generate bidirectional word vectors.Then CRF is used to calculate the joint probability to increase the dependence on neighboring word labels and optimize the whole sequence,so as to solve problems such as labeling bias.After passing through the sequence labeling model,we get a tag sequence that contains the correct and incorrect labels of the text.(3)The error correction methods of Chinese text are studied.The various strategies are adopted to make the correction of errors in view of the different types of errors,such as directly deleting the redundant words,adopting masked language model,confusion set matching and cluster search for wrong words,as well ass using sequence generation for missing words and disordered words to infer correct words.As for the masked language model,it is to replace the wrong words,missing words,errors or missing parts with [MASK],predict the correct position of wrong words and missing words,get several candidate words,and then select the most likely word by calculating the confusion degree of each candidate word.In accordance with the experimental results,the optical character recognition model based on Rep OCR proposed in this paper improves the recognition accuracy and extraction speed of picture text with different sizes.The Chinese error detection and correction model based on ERNIE and Bi LSTM-CRF sequence annotation has achieved good results compared with other models.