Research Projects

Many ancient manuscripts significant to Indian cultural heritage are preserved only in fragmented or deteriorated conditions, highlighting the critical need for restoration and preservation efforts. For example, the lost Indo-Buddhist Sanskrit manuscripts that survive only in Tibetan are often old and require expert analysis.  National initiatives like the National Mission for Manuscripts (NMM) and digital archives across India have driven efforts in manuscript preservation and Heritage science. However, preserving and visualizing historical documents fully requires advanced image restoration techniques. While deep generative models, such as GANs and diffusion models, have shown promise in image restoration, they remain underutilized in Heritage Science, especially for manuscripts with Indian historical and artistic value. Most generative models focus on general-purpose restoration and lack the specialized training needed for illustrated manuscripts containing both text and imagery. Therefore, a gap exists in research for models that can address the multimodal characteristics of culturally significant manuscripts. This project aims to tackle this gap by investigating deep learning models and providing solutions to restore and enhance images of culturally significant Indian documents. For simplicity, we refer to the restoration of digital manuscripts as manuscript restoration.  We plan to implement the following deep-learning models: Image Restoration Models (IRM), Vision Language Models (VLMs), and Neural Style Transfer (NST). We describe the research plan as follows. (1) Image Restoration Models (IRM) aim to improve degraded image quality. In the context of digital manuscript restoration, we plan the following tasks: Reconstruction of Damaged Sections, Noise Reduction, and Heritage Color Consistency. (2) Vision and Language Models (VLMs) combine visual and textual data for tasks like visual question answering and image captioning.  We plan to utilize VLMs for the following tasks: Script Classification, Segmentation of Script Elements, and Text-guided Image Editing. (3) Neural Style Transfer models (NST) apply the visual style of one image onto another, preserving content while transferring stylistic elements.  We plan NST models for the following tasks: Texture Preservation, Photorealistic Detailing, and Ink Clarity Enhancement. With the integration of IRM, VLMs, and NST, this project ensures that restorations are not only visually accurate but also culturally respectful. To validate the suggested models, we plan to include assessments from subject matter specialists.