Robust OCR for information extraction from retail flyers

Abstract

This report explores the augmentation of the RanLayNet research paper by addressing limitations in existing datasets for domain adaptation. Through a comprehensive analysis of Ran- LayNet, PubLayNet, and DocLayNet papers, deficiencies in dataset suitability for domain adaptation were identified. The study focused on leveraging a YOLOv8 model, fine-tuned using a subset of the PubLayNet dataset (1 of 13), and subsequently applied inferencing techniques on the multifaceted DocLayNet dataset, spanning Government Tenders, Laws and Regulations, Manuals, and Patents domains. This approach aimed to bridge gaps in dataset applicability and enhance document layout analysis for diverse domains. Results and implications of this methodology in the context of domain adaptation within document analysis are presented and discussed.