Nghiên cứu các phương pháp trích xuất thông tin trong ảnh tài liệu và ứng dụng

i - _ MSHV: 1970216 Ngày, tháng, 8.48.01.01 ages and application 22/02/2021 : 13/06/2021 N ii iii - - iv ABSTRACT Thanks to the development of information extraction models, it is possible to digitize and extract important information in document images quickly and efficiently In this thesis, we present an information extraction model using deep learning on graphs This is one of the newest and promising techniques for information extraction problems However, to complete a problem of extracting information in document images, we need to the following steps: Object detection, Text detection and Optical character recognition These are also things that we have researched and tested Regarding the proposed method, Object detection will be approached by Mask R-CNN model, Text detection will use text area detection method with CTPN model, Optical character recognition will adopt Tesseract OCR, which is developed by Google, and Information extraction will be approached by classifying text regions using a graph convolutional neural network model (GCN and GraphSAGE) For each process, we tested and evaluated our method with the collected dataset We tested the system on a dataset of English business cards of several companies Because the business card is a highly random, diverse, and useful form of data in practice Besides, we also conduct system evaluation in comparison to some commercial products on the market such as Abbyy, BizConnect v góp ý ... promising techniques for information extraction problems However, to complete a problem of extracting information in document images, we need to the following steps: Object detection, Text detection