Research on Multi-scale Printed Formula Recognition Methods

Mathematical formula recognition has wide applications in intelligent document processing, significantly reducing the time required to input mathematical formulas. We provide a deep neural network model based on an encoder-decoder architecture to handle the complexity of formula input. This model can convert mathematical formula image into the appropriate LaTeX markup sequences. The encoder, a convolutional neural network, converts the raw image into a collection of feature maps. The contribution of this paper is the introduction of the Feature Pyramid Transformer (FPT) into the encoder, combined with a residual network, to extract features at different scales, thereby improving the accuracy of the LaTeX sequence output. The attention-based Transformer decoder translates the encoder's output into a sequence of LaTeX symbols. The im2latex-100k dataset was used to train and evaluate this model. The experimental results confirm the effectiveness and superiority of this method.