Layout Analysis of Malayalam Palm Leaf Images with Fully Convolutional Networks

Palm leaf manuscripts have been used for centuries as a medium for preserving ancient knowledge and cultural heritage. To prevent their loss, digitization of these manuscripts has become crucial. Layout analysis specifically involves the localization and classification of various objects found in the document, such as text, tables, images, graphics, document boundaries, and background elements. However, layout analysis of palm leaf images presents significant challenges due to their irregular shapes and sizes, complex layouts, and varying degrees of deterioration. This research paper proposes the utilization of Fully Convolutional Networks (FCN) for the layout analysis of Malayalam palm leaf images. Multiple variations of FCN, including FCN-8, FCN-16, FCN-16, and FCN-8 with skipped connections, are explored to accurately extract and classify layout components such as background, text, and punch holes. The research uses a manually crafted masked dataset and combines manual research, deep learning, and image processing techniques. The results demonstrate the potential of FCN-based models for analyzing palm leaf manuscript layouts and contribute to their preservation and understanding. Significantly, FCN-8 yielded the best score, achieving a Mean Average Precision (MAP) of 0.79 at Intersection over Union (IOU). The paper provides an in-depth analysis of the methodology, dataset, and results, offering new avenues for research and preservation in this domain.