Intelligent Document Validation Using Computer Vision and Natural Language Processing
João Filipe Mendes Castilho · Portuguese National Funding Agency for Science, Research and Technology (RCAAP Project by FCT) · 2021
In recent years, more and more information has been produced that needs to be easily available and accessible.This need has led to an increasing digitalization of information in order to meet the demands of users.This process can be carried out through manual work, or by scanning files.The first option provides better results, since the information becomes "editable", however, it is a time-consuming, repetitive process and it's susceptible to failures.The second option is much faster, however, the information turns out to be in image format, which makes its management and organization difficult.Even if this second approach uses an Optical Character Recognition (OCR) tool, the information is not structured, which makes the research and organization process quite complex.CRITICAL Software (CSW) is developing a document analysis platform, which offers numerous services, including the document classification service and the information extraction and analysis service.The internship carried out at the CSW company, with the duration of an academic year, focuses on the extraction and analysis of information from documents.Its main objective is to implement a proof of concept of an algorithm for extracting and analyzing information from documents on the CSW platform, Intelligent Document Validation (IDV).The internship took place in three stages.In the first stage, research articles on state-of-the-art algorithms for document information analysis were explored and CloudScan[52], CUTIE [74], Chargrid [33] and BERTgrid [7] algorithms were selected.In the second stage, the algorithms selected in the previous step and respective variants were implemented in a total of 9 algorithms.In the third stage, an optimization analysis and performance comparison of the implemented algorithms was performed.These algorithms were tested on a data set composed of 1210 invoices and it was concluded that the algorithm with the best performance results from a variant that combines the Chargrid and BERTgrid algorithms, whose overall performance was 85.89% with the F1-Score metric .As a proof of concept, this version was successfully implemented on CSW's IDV platform.