Historical OCR Pipeline

Overview

This project implements an end-to-end Optical Character Recognition (OCR) system for transcribing 17th-century Spanish printed texts using a CRNN (Convolutional Recurrent Neural Network) architecture with CTC loss.

The system is trained on the RODRIGO corpus (~15,000 line images) and is designed to handle historical typography, ligatures, and diacritics found in early modern manuscripts.

---

Key Features

• End-to-end OCR pipeline using CNN + BiLSTM + CTC Loss
• No character-level annotations required (sequence learning via CTC)
• Handles historical Spanish text with diacritics and ligatures
• Beam search decoding with optional lexicon constraints
• LLM-based post-processing for error correction (Gemini / GPT-4o-mini)
• Evaluation using CER (Character Error Rate) and WER (Word Error Rate)

---

Architecture

CRNN (CNN + RNN + CTC)

• CNN (VGG-style): Extracts spatial features from text-line images
• BiLSTM: Models sequential dependencies in both directions
• CTC Loss: Enables alignment-free training for variable-length sequences

---

Design Choices

Architecture Choice: CRNN with CTC

A CRNN architecture is used because it enables sequence recognition from images without requiring character-level segmentation:

• CNN extracts column-wise visual features (strokes, shapes, typography)
• BiLSTM captures contextual dependencies and ligatures
• CTC loss aligns predictions with text sequences without explicit alignment

---

LLM Post-Correction

A lightweight LLM (Gemini / GPT-4o-mini) is used as a post-processing step:

• Corrects common OCR errors (e.g., rn → m, 0 → o, I → l)
• Improves linguistic coherence and spelling
• Does not replace the OCR model — only refines outputs

---

Evaluation Metrics

• CER (Character Error Rate):
  Measures character-level edit distance
  → Primary metric for OCR performance

• WER (Word Error Rate):
  Measures word-level errors
  → Sensitive to spacing and segmentation

Evaluation is performed:

• Before LLM correction
• After LLM correction

---

Text Region Detection

A morphology + contour-based approach is used to:

• Detect main text regions
• Remove marginalia, headers, and noise
• Ensure model processes only relevant content

---

Dataset

• RODRIGO OCR Dataset (Zenodo)
• ~15,000 line images from 16th–17th century Spanish texts

---

▶️ How to Run

pip install -r requirements.txt