ResearchActive Research Collaboration2024 — Present

Indic Palm Leaf Manuscript Analysis & Cultural Heritage Preservation

Role: Deep Learning Research Collaborator (IIT Tirupati)

Inter-institutional deep learning research collaboration with researchers at IIT Tirupati, developing multi-stage computer vision models to transcribe fragile historical manuscripts.

Verified Project Links & Repositories

3 Active Resources

GitHub Repository (Main)

Primary verified codebase, implementation, and commit history

Technical Documentation

Architecture diagrams, API references, and setup guide

Research Paper / Preprint

Formal academic write-up, methodology, and empirical evaluations

Indic Palm Leaf Manuscript Analysis & Cultural Heritage Preservation

Technology Stack & Libraries

PyTorchVision Transformers (ViT)OpenCVScikit-ImagePythonJupyterPandas

The Challenge / Problem

Context & Objectives

Millions of ancient palm-leaf manuscripts across Indian archives remain untranscribed and are actively decaying, with classical OCR failing completely due to non-uniform background degradation.

Architectural Solution

Engineered Approach

Designed an adaptive thresholding and Vision Transformer pipeline pre-trained on synthetic degraded glyphs, followed by fine-tuning on annotated historical Indic epigraphical samples.

System Architecture & Pipeline

Raw High-Res Manuscript Scan -> Dual-Stream U-Net for Biological Artifact Removal -> Contrastive Glyph Extraction -> Vision Transformer Classifier -> Text Stream Transcription.

Ancient Indic palm-leaf manuscripts suffer from severe background degradation, ink flaking, fungus discoloration, and physical fractures. This project formulates a robust binarization and character recognition architecture that separates ancient script glyphs from natural biological fiber noise.

Engineered Capabilities & Innovations

Multi-stage deep learning pipeline handling severe biological decay and text fissures

Adaptive contrast enhancement preserving delicate stylus stroke width

Segmentation of complex Indic ligatures and compound consonants

Inter-institutional validation with researchers at IIT Tirupati

Empirical Results & Benchmarks

Demonstrated an 18.4% improvement in Character Error Rate (CER) compared to baseline Sauvola and Otsu thresholding models on degraded sample sets.

Have Questions About This System?

I am always glad to discuss technical architecture, benchmarks, or potential collaboration.

Discuss This Project