Degraded Ancient Ashokan Brahmi Script Recognition
End-to-end OCR architecture and data generation pipeline targeting ICDAR (2027).
Project Overview
Leading a computer vision team alongside L. Chhetri, A. Anand, and G. Sarma to engineer a data generation and ETL pipeline using WGAN-GP to synthesize 20K+ Brahmi characters, applying physically-motivated degradation to construct a 150K sequence dataset.
Key Achievements
- Architecture Engineering: Developed an end-to-end OCR architecture integrating SimCLR self-supervised pretraining on a ResNet34 backbone with a BiLSTM-CTC decoder.
- Benchmark Establishment: Conducted a synthetic-to-real domain gap study, establishing the first severity-based Character Error Rate (CER) evaluation benchmark for ancient Indic scripts.
Role: ML/CV Lead Technologies Used: PyTorch, WGAN-GP, SimCLR, OCR Pipelines