MSNet: A Multi-scale Segmentation Network for Documents Layout Analysis

• Published in: Learning Technologies and Systems Vol. 12511; pp. 225 - 235
• Main Authors: Wang, Bo, Zhou, Ju, Zhang, Bailing
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Layout analysis; Multi-scale; Semantic segmentation
• ISBN: 303066905X; 9783030669058
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-66906-5_21

Layout analysis is often a crucial step in document image analysis and understanding. In this paper, we propose a deep learning-based layout analysis approach to identify and categorize the regions of interests in the scanned image of text document. Although semantic segmentation has been applied at pixel-level of document image for geometric layout analysis with much progress, many challenges remain with complex and heterogeneous documents which often have a sparse structure without closed boundaries and fine typologies with variable scales. We propose a multi-scale segmentation network, called MSNet, for high-resolution document image. The model is characterized by the enlarged receptive field size and multi-scale feature extraction. Experiments are conducted on a Chinese document dataset with satisfying performance.