Read Math Problems Using A Multi-stage Method

Problems recognition is an educational application.It can be utilized to digitize test papers,which reduces required space to store papers and the cost of maintenance and search for papers.And digitized test papers are helpful to develop future applications such as auto grading.Therefore,a system that can read problems is necessary.In this thesis,a multi-stage framework of reading math problems from test papers is proposed.The input images of the framework are transformed into images of characters through binarization,document layout analysis based on connected components,and a pixel-based approach of character segmentation.And then these images are fed into deep convolutional neural networks to classify and recognize the characters.To simplify the recognition task,images are first classified into Chinese and other elements through deep learning methods.And the other elements classified are fed into the models for recognition,while Chinese characters are recognized by Application Programming Interface(API).In the deep learning section,three classical models are applied in the research.The overall accuracy of the whole framework to read problems from 5 pages is 95.72%.In addition,two types of datasets are built in this thesis.One of them is collected from printed mathematical test papers and the other is generated by codes.