Handwritten Mathematical Formula Recognition Based On Codec And Self-attention Model

Handwritten mathematical formula recognition plays a crucial role in intelligent marking and is a research hotspot in character recognition.However,mathematical formula recognition still has many outstanding problems,such as character omissions,obfuscation,and exponential shifts.Therefore,it is important to be able to find a way to accurately identify handwritten mathematical formulas.In the modern teaching process,the handwritten mathematical formula recognition system of student test paper can effectively reduce the workload of teachers and improve work efficiency.This paper aims at the recognition of handwritten mathematical formulas,and designs a handwritten mathematical formula recognition system based on codec and self-attention mechanism model,which includes five modules: login registration,data input,image processing,test paper recognition and result output.Firstly,in the design process,the system recognition model is optimized,the encoder adopts a fully convolutional neural network,and the decoder adopts a long short-term memory neural network,which can input pictures of any size,perform feature extraction for handwritten mathematical formulas,and solve the gradient problem of explosion and disappearance;Secondly,the self-attention mechanism is introduced into the codec model to reduce the dependence on external information,help character segmentation reduce missed cuts,thus improving the accuracy of encoder segmentation,and prompting the decoder to focus on a specific part of the picture,so as to improve the recognition ability of handwriting mathematical recognition and reduce the recognition error rate.In this paper,the recognition rate and error rate of different recognition models are compared on the mathematical formula public test set CROHME 2014 and the self-made dataset.Among them,the recognition rate and error rate of the codec model used in this paper were 60.33% and 12.85% on the CROHME 2014 test set,and the recognition rate and error rate were 75.58% and 8.02% on the self-made dataset,respectively.Through the comparison of the difference between recognition rate and error rate,it can be seen that the codec model used in this paper has improved both recognition rate and error rate.