Information Extraction from Scanned Invoice Documents Using Deep Learning Methods
Ufuk Ilke Avei, Dionysis Goularas, Emin Erkan Korkmaz, Baris Deveci · 2024
In this paper, we explore innovative approaches in the field of information extraction from scanned invoice documents using deep learning methods. Our study makes significant contributions in three key areas. Firstly, we introduce a novel organizational method for labeling invoices, designed to enhance the efficiency and accuracy of data extraction. This lays a foundation for future research in this domain. Secondly, we break new ground by classifying a larger number of classes, 57 in total, far exceeding the typical 8–10 classes usually addressed in existing literature. This comprehensive classification enables a more detailed and nuanced understanding of invoice data. Lastly, we present our experimentation with various deep learning architectures, including Graph Convolutional Network (GCN), LayoutLMv1, and LayoutLMv3. Notably, our findings reveal promising, albeit preliminary results for Graph Convolutional Networks (GCN), an architecture that is not pre-trained, sug- gesting potential for further exploration and development in this area.