Data Synthesis for Document Layout Analysis

Layout analysis plays an important role in various document image processing tasks such as OCR and document understanding, and the methods based on deep learning have achieved significant achievements. In recent years, pre-training and transfer learning techniques have become a common practice in a variety of computer vision and natural language processing tasks. In this paper, we present an efficient approach of data synthesis for pretraining deep learning models in document layout analysis. The synthesized data is automatically annotated based on heuristic rules, and then applied to the PubLayNet pre-trained models. The models are fine-tuned with real document layout data. Three types of document elements are taken into account: text lines, tables, and figures/images. The experiments demonstrate that the pre-training model with synthesized data is very effective for transfer learning on different document domains.