Advanced document data extraction techniques to improve supply chain performance

Sharma, Vikash

Business; Systems science
July 2021

Thesis or dissertation


Rights
© 2021 Vikash Sharma. All rights reserved. No part of this publication may be reproduced without the written permission of the copyright holder.
Abstract

In this thesis, a novel machine learning technique to extract text-based information from scanned images has been developed. This information extraction is performed in the context of scanned invoices and bills used in financial transactions. These financial transactions contain a considerable amount of data that must be extracted, refined, and stored digitally before it can be used for analysis. Converting this data into a digital format is often a time-consuming process. Automation and data optimisation show promise as methods for reducing the time required and the cost of Supply Chain Management (SCM) processes, especially Supplier Invoice Management (SIM), Financial Supply Chain Management (FSCM) and Supply Chain procurement processes. This thesis uses a cross-disciplinary approach involving Computer Science and Operational Management to explore the benefit of automated invoice data extraction in business and its impact on SCM. The study adopts a multimethod approach based on empirical research, surveys, and interviews performed on selected companies.
The expert system developed in this thesis focuses on two distinct areas of research: Text/Object Detection and Text Extraction. For Text/Object Detection, the Faster R-CNN model was analysed. While this model yields outstanding results in terms of object detection, it is limited by poor performance when image quality is low. The Generative Adversarial Network (GAN) model is proposed in response to this limitation. The GAN model is a generator network that is implemented with the help of the Faster R-CNN model and a discriminator that relies on PatchGAN. The output of the GAN model is text data with bonding boxes. For text extraction from the bounding box, a novel data extraction framework consisting of various processes including XML processing in case of existing OCR engine, bounding box pre-processing, text clean up, OCR error correction, spell check, type check, pattern-based matching, and finally, a learning mechanism for automatizing future data extraction was designed. Whichever fields the system can extract successfully are provided in key-value format.
The efficiency of the proposed system was validated using existing datasets such as SROIE and VATI. Real-time data was validated using invoices that were collected by two companies that provide invoice automation services in various countries. Currently, these scanned invoices are sent to an OCR system such as OmniPage, Tesseract, or ABBYY FRE to extract text blocks and later, a rule-based engine is used to extract relevant data. While the system’s methodology is robust, the companies surveyed were not satisfied with its accuracy. Thus, they sought out new, optimized solutions. To confirm the results, the engines were used to return XML-based files with text and metadata identified. The output XML data was then fed into this new system for information extraction. This system uses the existing OCR engine and a novel, self-adaptive, learning-based OCR engine. This new engine is based on the GAN model for better text identification. Experiments were conducted on various invoice formats to further test and refine its extraction capabilities. For cost optimisation and the analysis of spend classification, additional data were provided by another company in London that holds expertise in reducing their clients' procurement costs. This data was fed into our system to get a deeper level of spend classification and categorisation. This helped the company to reduce its reliance on human effort and allowed for greater efficiency in comparison with the process of performing similar tasks manually using excel sheets and Business Intelligence (BI) tools.
The intention behind the development of this novel methodology was twofold. First, to test and develop a novel solution that does not depend on any specific OCR technology. Second, to increase the information extraction accuracy factor over that of existing methodologies. Finally, it evaluates the real-world need for the system and the impact it would have on SCM. This newly developed method is generic and can extract text from any given invoice, making it a valuable tool for optimizing SCM. In addition, the system uses a template-matching approach to ensure the quality of the extracted information.

Publisher
Business School, The University of Hull
Supervisor
Mishra, Nishikant; Merali, Yasmin
Qualification level
Doctoral
Qualification name
PhD
Language
English
Extent
5 MB
Identifier
hull:18527
QR Code