A study on handwritten parcel delivery invoice understanding model

다국어 초록 (Multilingual Abstract)

Optical character recognition (OCR) technology is a field of continuous research in which text images are stored or utilized as data. However, OCR technology alone has limitations in classifying and recognizing attribute values such as address, name, and phone number expressed in text in a semi-structured form, such as a parcel delivery invoice (PDI) written by hand. Therefore, in this study, we propose a handwritten parcel delivery invoice understanding (HPDIU) model for automated parcel delivery reception. The proposed HPDIU model consists of two steps: region detection of the parcel delivery invoice (RD-PDI) and information extraction from the PDI (IE-PDI). The RD-PDI, which is the first step, minimizes the resolution adjustment by detecting only the necessary area, including the sender and recipient information in the image. The second step, IE-PDI, consists of an end-to-end framework without OCR technology using a document understanding transformer and integrates the process of character detection, recognition, and understanding.
In other words, the proposed model can solve the limitations of the OCR technology because it can integrate the process of classifying and recognizing according to attributes. To prove the validity of the proposed model, we used 500 handwritten PDI datasets to evaluate the accuracy of the character units according to the attributes of address, name, and phone number. As a result of the evaluation, a total average of 91.67% in units of letters proved the superiority of the proposed HPDIU model.

참고문헌 (Reference)

1 Z. Liu, "Swin transformer: Hierarchical vision transformer using shifted windows" 9992-10002, 2021

2 T. T. H. Nguyen, "Survey of post-OCR processing approaches" 54 (54): 1-37, 2022

3 D. Wang, "Rout-ing and scheduling for hybrid truck-drone collaborative par-cel delivery with independent and truck-carried drones" 6 (6): 10483-10495, 2019

4 B. Shi, "Robust scene text recognition with automatic rectification" 4168-4176, 2016

5 B. P. Majumder, "Representation learning for information extrac-tion from form-like documents" 6495-6504, 2020

6 W. Hwang, "Post-OCR parsing: Building simple and robust parser via bio tagging" 2019

7 G. Kim, "OCR-free document under-standing transformer" 498-517, 2022

8 S. Rijhwani, "OCR post correction for endangered language texts" 5931-5942, 2020

9 A. Buades, "Non-local means de-noising" 1 : 208-212, 2011

10 A. Radford, "Improving language understanding by generative pre-training" 2018

11 R. Tito, "ICDAR 2021 competition on document visual question answering" 635-649, 2021

12 W. Hwang, "Findings of the Association for Computational Linguistics" 330-343, 2021

13 E. Tallyn, "Enacting the last mile: Experiences of smart con-tracts in courier deliveries" 1-14, 2021

14 M. Mathew, "DocVQA: A dataset for VQA on document images" 2199-2208, 2021

15 M. Z. Afzal, "Deepdocclassifier: Document classification with deep convolutional neural network" 1111-1115, 2015

16 W. Hwang, "Cost-ef-fective end-to-end information extraction for semi-struc-tured document images" 3375-3383, 2021

17 L. Kang, "Con-volutional neural networks for document image classifica-tion" 3168-3172, 2014

18 D. Lewis, "Building a test collection for complex docu-ment information processing" 665-666, 2006

19 J. Devlin, "Bert: Pre-training of deep bidirectional transformers for language un-derstanding"

20 M. Lewis, "Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension" 7871-7880, 2020

21 Q. Duong, "An unsuper-vised method for OCR post-correction and spelling normal-isation for Finnish" 240-248, 2020

22 B. Shi, "An end-to-end trainable neural network for image-based sequence recognition and its ap-plication to scene text recognition" 39 (39): 2298-2304, 2017

23 R. Schaefer, "A two-step approach for automatic OCR postcorrection" 52-57, 2020

24 N. Islam, "A survey on optical char-acter recognition system"