Design Of Mathematical Formula Recognition System Based On Convolutional Neural Network And Attention Mechanism

The development of artificial intelligence technology has promoted the construction of a smart education ecosystem and the reform of traditional education network and intelligence.At present,there have been many intelligent applications such as intelligent marking and self-built question banks.Among them,the recognition of mathematical formulas with a two-dimensional structure has played a pivotal role.Therefore,a system that can accurately recognize mathematical formulas is of great significance.In this thesis,a recognition model combining visual attention mechanism and encoding-decoding framework are used to build a middle school mathematical formula recognition system.The system consists of 3 parts,which are input module,recognition module,and output module,which can support the recognition of two types of fonts,printed and handwritten.In the process of constructing the recognition system,a lot of innovative work is involved.First,a network model with higher comprehensive performance is proposed as the encoder,and the decoder is composed of GRU and attention mechanism.Secondly,building dataset containing 100,000 print's math formulas.And,a mathematical logic grammar correction library was established to check and correct the predicted La Te X characters.In the end,the system achieved a sequence recognition accuracy of 98.5% on a dataset composed of 3000 middle school mathematical formulas.The average recognition speed on the CPU is 0.41 s and the average recognition speed on the GPU is 0.22 s,which are better than the current formula recognition system.In addition,this thesis also analyzes and compares the advantages and disadvantages between the traditional formula recognition method and the deep learning recognition method.The recognition accuracy of 51.6% on the public test set CROHME 2014 is achieved,which is better than the traditional formula recognition model and most deep learning models.In order to support and adapt to more application scenarios,it was deployed on the artificial intelligence embedded platform Jetson TX2.