An end-to-end image text entity extraction method

Abstract

This invention discloses an end-to-end image text entity extraction method, comprising: Step 1: acquiring the image to be recognized, and extracting feature vectors from the image through the convolutional layer in a preset network model; Step 2: the encoding layer in the preset network model performs feature fusion on the extracted feature vectors from the image, fusing global knowledge into each feature vector to obtain global information, adding residuals, and then compressing the dimension through the feedforward layer in BERT to obtain the output feature dimension. The features output by BERT are then processed by a deep bidirectional LSTM neural network to extract text sequence information; Step 3: the decoding layer in the preset network model decodes the text sequence information in the encoding layer to obtain the final recognition result. This invention utilizes a network architecture of convolutional neural network + BERT model + deep bidirectional LSTM neural network + CRF model, which improves the recognition accuracy and robustness of the model.

Images