End-to-End Recognition Algorithm for Degraded Invoice Text.

Jiahao Wang,Liping Zhu,Xin Yi,Tong Lu

DOI: https://doi.org/10.1109/CISP-BMEI60920.2023.10373339

2023-01-01

Abstract:With the improvement of social informatization and computer technology, the effective realization of electronic management and application of a large number of paper invoices has become a new challenge in industrial practice. Text detection and recognition technology is an important means to realize the structured information extraction of paper invoices. However, the straight lines, symbols and other elements in paper invoices easily interfere with text detection, and the drawings and characters are prone to adhesion, intersection, overlap and other degradation during the electronic process. In addition, the detection and recognition process of invoice text is relatively relevant, and the existing methods based on deep network cannot be applied to the detection and recognition of it. Based on this, this paper proposes a new end-to-end recognition framework Invoice Text Recognizer (ITR) for complex invoice texts. Specifically, the ITR uses the conversion encoder with self attention mechanism as the detector, and uses the fusion recognition module to coordinate the detection and recognition of the two task features. It does not require additional correction modules, nor does it need to make character level annotation on any shaped text. It can directly guide the text location through the loss in the recognition phase. The experimental results on three open text datasets (ICDAR2015, Total Text, CTW1500) and the paper-based invoices of a company prove that the proposed ITR has a high detection and recognition rate.