AutoDocSegmenter: A Geometric Approach Towards Self-Supervised Document Segmentation

Abstract

Document segmentation, the process of dividing a document into coherent and significant regions, plays a crucial role for diverse applications that require parsing, retrieval, and categorization. However, most existing methods rely on supervised learning, which requires large-scale labeled datasets that are costly and time-consuming to obtain. In this work, we propose a novel self-supervised framework for document segmentation that does not require labeled data. Our framework consists of two components: (1) an unsupervised isothetic covers based pseudo mask generator which approximately segments document objects, and (2) an encoder-decoder network that learns to refine the pseudo masks and segments the document objects accurately. Our approach can handle diverse and intricate document layouts by leveraging the rich information from unlabeled datasets. We demonstrate the effectiveness of our approach on several benchmarks, where it outperforms state-of-the-art document segmentation methods. Our code is available at https://github.com/ankitachatterjee94/AutoDocSegmenter