A Top-down Attention Based Approach For Printed Mathematical Expressions Recognition

Mathematical expressions are widely present in various types of literature and are an essential part of scientific and technical documents.The need to identify expressions in PDF docu-ments has increased recently,but the recognition of mathematical expressions is far more complicated than the recognition of paragraphs.Since mathematical expressions have unique structure,it is not easy to deal with these expressions by using traditional natural language search systems.Mathematical information retrieval is a vital research hotspot in the field of mathematical expressions recognition.Off-line printed mathematical expression recognition is different from traditional printed character recognition.The mathematical expression is a symbol system with a two-dimensional layout.To correctly recognize mathematical expres-sions,not only must correctly recognize mathematical symbols,but also accurately identify the structural relationships between the symbols.It is a challenging task in the field.In this thesis,the critical techniques of mathematical expression recognition are intensely studied.A solution based on Encoder-Decoder language model for the recognition of mathematical expression is proposed.In addition,Attention Mechanism is introduced into the model to promote the performance.With this model,we can convert the printed mathematical expression image to latex sequence.Unlike the traditional methods,the training process of the model is end-to-end which means we skip the heavy pipeline steps such as symbol segmentation.We evaluate the model on IM2LATEX-100K dataset and CROHME 2016 and the match score id 54.90%and 38.74%respectively.The model achieved good results under limited experimental conditions and time.