A Novel Text Recognition Method Based on Counting Focusing Model

Abstract

The present invention relates to a novel text recognition method based on a counting focus model, the counting focus model includes an encoder and a decoder, and the recognition method includes the following steps: S1. Using an encoder based on a convolutional neural network to process an input image The high-level features are extracted to obtain the high-level feature map; S2. The decoder based on the long-term short-term memory network and the focusing mechanism decodes the characters from left to right in sequence from the high-level feature map.

Description

technical field [0001] The invention belongs to the field of optical character recognition, and more specifically relates to a novel text recognition method based on a counting focus model. Background technique [0002] OCR single-line text recognition is the process of recognizing the text content of an input image containing a single-line text. One of the mainstream models currently used on this task is the attention / focus model (Attention Model), and its recognition process is: [0003] 1) First use the convolutional neural network (CNN) to extract the high-level feature map (feature map) of the input image; [0004] 2) Use the long-term short-term memory network (LSTM) to "attend" the high-level feature map multiple times, and calculate the attention weights (attention weights); [0005] 3) Use the focus weight to perform weighted average of the high-level feature maps, and predict the text characters that need to be output at the current step (step) according to the o...

AI Technical Summary

Problems solved by technology

[0009] The present invention solves the technical defect that the prior art does not assume the relative position of successive focus positions, and needs to allow the model to learn to focus from left to right or from top to bottom during the training process, resulting in a complicated training process, and to calculate the focus weight The design of the module is too complicated for technical defects, and a new text recognition method based on focus weight is provided.

Images