Sri Lankan Institute of Information Technology

Silent Script

An AI-powered pipeline to digitize, restore, and translate ancient Sri Lankan inscriptions from degraded stone surfaces.

Developed by: Lahiru Bandara, Amadhi Hansani, Senod Mesandu, Chamodya Handapangoda

Supervised by: Thamali Dassanayake, Samadhi Rathnayake

Read Paper

Project Overview

Silent Script introduces a novel deep-learning architecture designed to automate the epigraphical analysis of ancient Sri Lankan inscriptions. By combining image enhancement, instance segmentation, and neural machine translation, the system successfully digitizes heavily degraded stone texts where traditional OCR fails.

Overall Accuracy (%)

10-20

Seconds Processing Time

Restoration Improvement (%)

Improved from <20%

The Challenge

Preserving the rich history embedded in ancient stone inscriptions is a race against time. Natural weathering and human interference have left these artifacts severely degraded, making traditional preservation methods insufficient.

No Existing OCR: Traditional Optical Character Recognition systems fail completely on complex, degraded stone surfaces.
Time Consuming: Manual transcription and translation by expert epigraphists takes 2-4 hours per inscription.
Severe Degradation: Centures of weathering make characters barely visible to the naked eye.

System Architecture

Stage 1: Image Enhancement

Advanced pre-processing using deep learning to remove noise, correct lighting, and highlight faint carvings on stone.

Stage 2: Damage Restoration

Generative models reconstruct missing or heavily weathered parts of characters, boosting readability from <20% to >90%.

Stage 3: Line & Word Segmentation

Instance segmentation algorithms isolate individual lines and words from the continuous, often irregular ancient script.

Stage 4: Character Segmentation

Precise extraction of individual characters, handling overlapping and joined ancient letters.

Stage 5: Character Recognition

A custom CNN classifies ancient characters with 94% accuracy, mapping them to modern digital equivalents.

Stage 6: Translation

A specialized Neural Machine Translation model converts the recognized ancient syntax into modern readable text.

Key Components

Character Recognition

Custom convolutional neural network trained on a newly curated dataset of ancient scripts to achieve high-precision classification.

Lahiru Bandara

Image Enhancement

Generative Adversarial Networks (GANs) specialized in texture filtering and contrast enhancement for lithic surfaces.

Chamodya Handapangoda

Word Segmentation & Translation

An integrated pipeline using spacing detection and Transformer-based NLP models to decode and translate ancient syntax.

Amadhi Hansani

Damage Restoration

Inpainting algorithms utilizing contextual awareness to reconstruct eroded or broken characters based on surrounding text.

Senod Mesandu

Experimental Results

System Accuracy Metrics

Character Recognition Accuracy 94%

Segmentation Precision 91%

Translation BLEU Score 88%

Table I: Model Comparison for Character Recognition

Model	Precision	Recall	F1-Score	Accuracy
Baseline CNN	0.82	0.80	0.81	81.5%
ResNet-50	0.89	0.88	0.88	88.2%
Proposed Arch (Ours)	0.95	0.93	0.94	94.1%