Business Documents Data Recognition and Table Extraction Using Deep Learning Techniques

As trade and business increasingly occur through digital platforms, the need to process physical business documents, such as invoices and receipts, and store them digitally grows. Digitalization has many benefits, such as safe storage, easy search and recovery, faster communication, and reliability. Although significant progress has been made in character recognition and extraction, an end-to-end framework tailored to processing specific business documents is not yet available as open-source software. This paper presents a deep learning approach for automating the data extraction process from business documents. The solution proposed uses two separate neural networks, YOLOv5, for identifying important information in documents and a transformer model to extract information from tables in the document. The method was evaluated on a dataset containing real-world invoices. The YOLO model performed satisfactorily, giving an mAP value of 84.36%. The table extraction model gave an mAP value of 80.92%. Once the model is trained, it can execute independently without intervention, demonstrating the effectiveness of reducing manual data entry requirements and errors.