MC-OCR Challenge 2021: An end-to-end recognition framework for Vietnamese Receipts

Abstract

Recognizing text from receipts is a significant step in automating office processes for many fields such as finance and accounting. MC-OCR Challenge has formed this problem into two tasks (1) evaluating the quality, and (2) recognizing required fields of the captured receipt. Our proposed framework is based on three key components: preprocessing with receipt detection using Faster R-CNN, alignment based on the angle and direction of rotation; estimate the receipt image quality score in task 1 using EfficientNet-B4 which has been retrained using transfer learning; while PAN is for text detection and VietOCR1 for text recognition. In the final round, our systems have achieved the best result in task 1 (0.1 RMSE) and a comparable result with other teams (0.3 CER) in task 2 which demonstrated the effectiveness of the proposed method.