Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
215 lines (139 loc) · 4.62 KB

ARCHITECTURE.md

File metadata and controls

215 lines (139 loc) · 4.62 KB

RVisionT — Architecture Overview

This document describes the architectural design and engineering decisions behind RVisionT, an engineering-oriented OCR pipeline.

The goal of the architecture is not maximum OCR accuracy, but robustness, stability, and controlled search complexity when working with unstable OCR engines.

High-Level Design

RVisionT implements a two-phase OCR pipeline with adaptive preprocessing, heuristic scoring, and explicit failure handling.

Core principles:

  • reduce OCR search space
  • avoid unbounded brute-force
  • isolate unstable operations
  • prefer stability over raw accuracy

Pipeline Overview

Input Image
     │
     ▼
Preprocessing + Normalization
     │
     ▼
Phase A — Orientation & Baseline Search
     │
     ▼
Best Orientation Selected
     │
     ▼
Phase B — Focused OCR Refinement
     │
     ▼
Optional ROI-Based OCR Fallback
     │
     ▼
Post-processing & Cleanup
     │
     ▼
Final Text Output

Phase A — Orientation & Baseline Search

Purpose

Phase A determines:

  • correct image orientation
  • baseline OCR configuration

This phase prioritizes speed and safety.

Strategy

Phase A iterates over:

  • image rotations (0°, 90°, 270°)
  • a limited set of Tesseract PSM modes
  • language configurations (rus, eng, rus+eng)

Key design choices:

  • uses image_to_string only (lower memory usage)
  • avoids aggressive thresholding
  • applies early stopping once a confident result is found

Output

Phase A produces:

  • best rotation
  • baseline OCR result
  • metadata describing the chosen configuration

Phase B — Focused OCR Refinement

Purpose

Phase B refines OCR quality using the best orientation selected in Phase A.

This phase trades speed for quality but within a controlled search space.

Strategy

  • runs only on a single orientation
  • evaluates multiple preprocessing variants
  • selectively uses image_to_data to estimate confidence
  • computes a composite heuristic score for each candidate

Scoring factors include:

  • text length
  • average confidence
  • character distribution
  • noise and garbage penalties

ROI-Based OCR Fallback

Motivation

Full-frame OCR often performs poorly on:

  • meme-style images
  • captions located at top/bottom regions

Approach

A lightweight text-band detection heuristic is applied:

  • horizontal density analysis
  • band merging and filtering
  • top/bottom caption heuristics

Each detected region is processed independently using Phase B logic, and results are merged top-to-bottom.

This fallback is applied only when layout heuristics indicate captions.

Preprocessing Architecture

Preprocessing is variant-based, not fixed.

Examples:

  • grayscale / inverted grayscale
  • unsharp masking
  • adaptive thresholding (Phase B only)
  • bilateral filtering

Each preprocessing variant is treated as a candidate and scored independently.

This avoids committing to a single preprocessing strategy.

Heuristic Scoring System

RVisionT uses a fully heuristic quality scoring system due to the absence of labeled training data.

Key scoring components:

  • character composition (letters vs noise)
  • script consistency (Cyrillic vs Latin)
  • word structure analysis
  • confidence weighting (when available)
  • penalties for OCR artifacts

This system prioritizes readable text over longer but noisy output.

Stability & Failure Handling

OCR engines such as Tesseract are treated as unstable dependencies.

Explicit safeguards include:

  • downscaling large images before OCR
  • skipping unsafe image sizes
  • limiting image_to_data usage
  • strict timeouts on OCR calls
  • catching and isolating OCR engine failures

At no point does a single OCR failure terminate the pipeline.

Architectural Trade-offs

Chosen Trade-offs

  • Heuristics instead of ML models
  • Controlled search instead of exhaustive brute-force
  • Stability over maximum accuracy
  • Readability over raw text length

Intentionally Not Addressed

  • deep learning text detection
  • production-grade APIs
  • large-scale OCR benchmarking
  • language model post-correction

These are considered future extensions, not architectural goals.

Extensibility

The architecture allows future additions such as:

  • learned scoring functions
  • external text detectors (EAST, CRAFT)
  • language-specific post-processing modules
  • GUI or service wrappers

Summary

RVisionT demonstrates a robust OCR pipeline architecture designed around real-world OCR instability and noisy input data.

The project emphasizes:

  • systematic engineering thinking
  • explicit failure control
  • adaptive processing pipelines
  • informed architectural trade-offs