Form extractor: information extraction from digital documents

In the digital age, businesses are increasingly focused on automating the processing of unstructured documents, such as invoices, contracts, and other text-heavy records. This thesis, completed during my internship at Efficiento, explores the development of an advanced solution for intelligent information extraction from digital documents. The project leverages state-of-the-art techniques in Computer Vision (CV) and Natural Language Processing (NLP) to automate the extraction of key data that would traditionally require manual input. By utilizing the MMOCR (Multimodal OCR) framework, the solution effectively handles various document formats, enabling accurate detection, recognition, and extraction of critical information. The system integrates DBNet for text detection, ABINet for text recognition, and the Spatial Dual-Modality Graph Reasoning (SDMG-R) model to enhance information extraction by combining both visual and textual data. Tested on a combination of public and proprietary datasets, the solution achieved high accuracy in extracting essential information. The results demonstrate a significant improvement in both efficiency and precision compared to manual methods, offering a scalable solution adaptable across a wide range of industries.