Deep Learning-based End-to-End Address Recognition Solution on Chinese Courier Order Forms

Abstract

The courier industry in China has grown quickly due to the rise of online shopping. However, courier notes can unavoidably become smudged or damaged during the delivery process, making it difficult to read the printed Chinese address information or recognize the barcodes. To solve this problem, this paper proposes an end-to-end solution to recognize damaged Chinese addresses: the CRNN model is trained for address recognition for damaged Chinese courier orders using a large Chinese address dataset generated via data augmentation and manual collection. And an address association algorithm is proposed to reduce the recognition errors at the provincial and municipal levels of the addresses. By applying this algorithm, the final accuracy is increased by 2% to 98.7%.