Deep Layout Analysis of Multi-lingual and Composite Documents

Abstract

It is crucial to accurately analyze the layout to convert document images to high-quality text. With the emergence of publicly available, large ground-truth datasets, deep-learning models have demonstrated their effectiveness in detecting and segmenting document layouts. This study presents a deep learning technique for document structure analysis, an important stage in the optical character recognition (OCR) system. Our method employs the YOLOv7 (Only Look Once version 7) model, a highly efficient and precise object detection model trained on the DocLayNet database. The trained YOLOv7 model quickly and efficiently identified and categorized different document components, such as caption, list item, text, table, section header, and picture. Regarding accuracy and efficiency, our evaluation demonstrates that the suggested method beats existing strategies, with strong generalization ability for diverse document layouts, text styles, and scripts.