Mathematical formula symbol identification method

Abstract

The invention discloses a mathematical formula symbol identification method, which comprises the following steps of: processing and identifying mathematical formula symbols by using an implicit segmentation model according to a mathematical formula to be identified, wherein in the process of processing and identifying the implicit segmentation model, the real-time position judgment result of the direction judgment network on the mathematical formula symbol is used as the input of the implicit segmentation model processing to be processed, and the implicit segmentation model identifies the mathematical formula symbol. And a direction judgment network is added to enhance the encoding and decoding capabilities in a two-dimensional space. By adopting the method, the accuracy of online handwriting, offline handwriting and optical character mathematical formula recognition is greatly improved.

Description

technical field [0001] The invention belongs to the technical field of pattern recognition and artificial intelligence, and in particular relates to a mathematical formula symbol recognition method. Background technique [0002] With the popularization of artificial intelligence and a large number of mobile terminals, using the method of directly inputting data by handwriting has become the most natural way for users to interact with computers. Mathematical formulas are two-dimensional documents, but at the same time, symbol information in one-dimensional space and space information between symbols need to be transmitted. Different symbols and spatial positions express different meanings. Therefore, how to accurately and quickly recognize the symbols in handwritten mathematical formulas and the positional relationship between the symbols is a problem that needs to be resolved. [0003] On-line Handwritten Character Recognition (HWR) is a technology to recognize handwritten...

AI Technical Summary

Problems solved by technology

In this method, although the problem of cutting and dynamic programming decoding is solved by RNN, which reaches the end-to-end level, because RNN (RNN, LSTM, GRU, etc.) internally encodes the time axis direction, it adopts the same method as above. The way of full connection at one time, although there is a gate mechanism in the structure similar to LSTM or GRU to control the long-short-term (Long Short-Term) associative memory ability, but due to the limitation of full connection and easy over-fitting, the computational complexity is high , Weak self-encoding ability and other reasons, the long-term associative memory ability decays seriously over time, resulting in a bottleneck in overall performance and effect

At the same time, if faced with the identification of complex structures such as mathematical formulas, the RNN part cannot take into account the correlation between the horizontal direction and the vertical direction, so it cannot effectively identify complex spatial structures

Images