Research On Printed Mathematical Formula Symbols Recognition Method Based On Deep Learning

The recognition of printed mathematical formula is an important research subject in the field of optical character recognition.It aims at turning printed mathematical formula which is input by image into the editable symbols, and then realizes formula reuse. However, because of the complex mathematical formula structure, many different types of mathematical symbols as well as diverse font and size, the recognition effects are not ideal, there needs to be a more effective recognition method. This paper will apply conventional neural network of the deep learning to mathematical symbol recognition, aiming to overcome the disadvantages of feature extraction by hand lying in traditional symbol recognition methods. By constructing, the deep network model with several hidden layers can learn more useful features by using a lot of training data, thus raising the accuracy of the formula symbols classification.This paper firstly researches on the key technology of mathematical formula recognition,status and the convolution neural network theory, on the basis of CNN, the convolution neural network structure for formula is established. The network is composed of one input layer, one output layer, two alternating layers of convolution and pooling. The model firstly receives two-dimensional images of the symbols which are from formula location and segmentation,and then uses the size of a 5×5 kernel to extract features. In the convolution layer, the traditional Sigmoid and hyperbolic tangent activation functions are replaced by ReLU activation function, which is closer to biological neural and improves the convergence speed of training models as well as solving the problem of gradient disappeared. In addition to the extraction of the convolution kernel feature, the maximum pooling method is used in the pooling layer with 2×2, which not only reduces the dimension of features but also maintains characteristics at the same time, thus reducing the parameters calculation. This method combining convolution and pooling makes features have rotation invariance. Meanwhile, the Dropout technology of full connected layer reduces the over fitting phenomenon of the network. Finally, the full connected layer links to classify symbols. Aiming at the problem of slow speed of convolution neural network training, this paper uses GPU programming method which is based on CUDA and has significantly enhanced the training speed.To verify the validity of the proposed algorithm, this paper also applies VS2010 to design and implement a printed mathematical formula recognition system. Taking document image containing formula as the input, after layout analysis, formula image pre-processing,recognition of formula symbols and formula structure analysis, it will output the result.Through a number of experiments and comparison, the view of the paper put forward concerning formula symbols, the average recognition rate can reach 99%, which is higher than the existing recognition methods, and the result can satisfy the need of practical application.