Deep Learning for Optical Character Recognition and Its Application to VAT Invoice Recognition

摘要

Optical character recognition (OCR) is considered as one of long-term and hot research topics due to the fact that OCR technique can change the documents from paper to computer-readable format by consistently growing. However, the recognition accuracy of current OCR technique is required to improve some special applications such as in reimbursement of value-added tax (VAT) invoices. This paper proposes two OCR techniques by using deep convolutional neural network (CNN) and residual network (ResNet), respectively. According to our test dataset, the formerly proposed techniques can reach up to 97.08%, while the latter can increase to 99.38%.