MC-OCR Challenge 2021: Deep Learning Approach for Vietnamese Receipts OCR

Abstract

Receipts OCR has made a significant improvement on accounting, which has attracted much attention of the research community in the field of computer vision as well as natural language processing. In this paper, we solve the problem of extracting pieces of information on Vietnamese receipts including seller, address, timestamp, and total cost. We divided this into two problems: detecting locations of information and using an OCR model to recognize texts. In this paper, we propose a pipeline that employs Faster R-CNN as an information location detector and training a Transformer model for text recognition. Through experiments, we achieved CER 32.19%, which is 9.65% higher than previous method CRNN, while pointing out the remaining statements and challenges of this problem.