RVisionT

Engineering-Oriented OCR Pipeline (Two-Phase, Stable)

RVisionT is an experimental Optical Character Recognition (OCR) project built with Python, OpenCV, and Tesseract OCR.

The project focuses on engineering robustness, stability, and adaptive OCR configuration rather than maximum accuracy or end-to-end deep learning models.

What this project demonstrates

• Design of a multi-stage OCR pipeline
• Robust handling of unstable third-party OCR engines (Tesseract)
• Heuristic-based quality scoring and early stopping
• Trade-offs between accuracy, performance, and stability
• Computer Vision preprocessing without deep learning

What this project is NOT

• Not a production-ready OCR system
• Not a deep learning OCR framework
• Not optimized for maximum recognition accuracy

This project is intentionally positioned as a research and engineering showcase.

Problem Statement

Vanilla Tesseract OCR often fails on:

• meme-style images
• mixed-language content (Russian + English)
• large screenshots
• noisy or stylized text

Common issues include:

• unstable results
• excessive garbage output
• memory crashes (std::bad_alloc)

Solution Overview

RVisionT implements a Two-Phase Stable OCR Pipeline designed to:

• reduce search space
• prevent crashes
• automatically select the best OCR configuration

System Architecture

Phase A — Orientation & Baseline Search

• Iterates over:
  • image rotations (0°, 90°, 180°, 270°)
  • Tesseract PSM modes
  • language configurations (rus, eng, rus+eng)
• Uses fast OCR (image_to_string)
• Determines the best orientation and baseline configuration
• Applies early stopping to limit computation

Phase B — Focused Refinement

• Runs only on the best orientation selected in Phase A
• Uses image_to_data for detailed OCR analysis
• Selects the final result using a composite heuristic score:
  • text length
  • average confidence
  • noise and garbage penalties

This approach:

• significantly reduces OCR calls
• avoids Tesseract memory overflows
• improves overall pipeline stability

Image Preprocessing

Multiple preprocessing variants are evaluated automatically:

• grayscale / inverted grayscale
• adaptive thresholding
• sharpening
• bilateral filtering

Each variant is scored independently, and the best result is selected dynamically.

Text Post-processing

General Normalization

• removal of garbage characters
• whitespace normalization
• correction of common OCR artifacts

Caption-Oriented Cleanup

A specialized sanitizer is applied only for caption-like text:

• fixes merged or split words
• restores missing punctuation
• removes outline-related noise common in meme images

Stability & Error Handling

Explicit safeguards include:

• limited use of image_to_data
• candidate pre-selection
• downscaling of large images
• graceful fallback on Tesseract errors
• no unbounded brute-force search

As a result, the pipeline remains stable even on complex or noisy images.

How to Run

Requirements

• Python 3.9+
• Tesseract OCR (installed and available in PATH)

Install dependencies:

pip install -r requirements.txt

Run:

python src/main.py path/to/image.jpg

Example Output

=== RVisionT Universal OCR (Two-Phase Stable) ===
Text:
It was the best of times, it was the worst of times,
it was the age of wisdom, it was the age of foolishness...

Confidence: 95.9

Known Limitations

• Heuristic-based scoring (no labeled dataset)
• Limited language support (Russian / English)
• OCR speed may be slow on large images
• ROI detection is heuristic, not layout-aware

Educational Value

This project demonstrates:

• practical Computer Vision techniques
• engineering strategies for unstable libraries
• search-space optimization
• noisy data handling
• systematic pipeline design

It is suitable as:

• a research or coursework project
• a foundation for future OCR extensions
• a portfolio example of engineering thinking

Author

XCON | RX TG: @End1essspace GitHub: End1essspace

Developed for educational and research purposes. All logic is implemented manually without using ready-made end-to-end OCR frameworks.