abstract-ocr

Part of the Abstract Media Intelligence Platform

This module provides layout-aware OCR as part of a larger media processing system.

abstract_ocr focuses on extraction:

• multi-engine OCR (Tesseract / EasyOCR / PaddleOCR)
• column detection and region segmentation
• structured, position-aware text output

Full system: https://github.com/AbstractEndeavors/abstract-media-intelligence

---

abstract_ocr / layout_ocr — Layout-Aware OCR Pipeline

A structured OCR pipeline designed for layout-aware text extraction from complex documents, combining preprocessing, column detection, region classification, and ordered OCR assembly.

Built to handle:

• multi-column PDFs
• mixed-content layouts (text, figures, captions)
• noisy or scanned documents
• large-scale document ingestion pipelines

---

🔹 What This System Is

This is not a simple OCR wrapper — it is a typed, multi-stage processing pipeline:

• transforms raw images into structured page representations
• detects document layout (columns, headers, regions)
• classifies content blocks (text, figures, captions)
• applies OCR at the region level
• reconstructs output in correct reading order

The system is designed for deterministic, reproducible extraction rather than heuristic text scraping.

---

Pipeline Overview

PDF Input
    ↓
Slice / Decompose (images + text per page)
    ↓
OCR + Text Extraction (layout-aware engines)
    ↓
Metadata Generation
    ├─ summaries
    ├─ keywords
    └─ descriptions
    ↓
Manifest Creation (per-page + per-document)
    ↓
HTML Generation
    ├─ PDF viewer pages
    └─ gallery index pages
    ↓
Static Site Output (SEO-ready)

flowchart TD
    A[Input Image / Page Image]
    B[Preprocess\nDenoise + Binarize]
    C[Layout Detection\nColumns + Header Cutoff]
    D[Region Classification\nText / Figure / Caption]
    E[Region OCR\nCrop + Tesseract]
    F[Fallback OCR\nColumn-level OCR]
    G[Reading Order Assembly]
    H[Structured OCRResult\nBlocks + Raw Text + Layout]

    A --> B --> C --> D --> E --> G --> H
    D -->|No usable regions| F --> G

---

🔹 Core Capabilities

• Layout Detection

  • Column detection via vertical projection valleys
  • Header segmentation via density scanning
  • Multi-column classification (single / dual / mixed)

• Region Classification

  • Connected-component analysis
  • Density-based classification (text vs figure vs caption)
  • Column-aware region assignment

• Region-Level OCR

  • OCR applied per detected block (not full-page)
  • Adaptive Tesseract configuration by region type
  • Automatic fallback to column-level OCR when detection fails

• Reading Order Reconstruction

  • Column-aware ordering
  • Top-to-bottom sequencing within columns
  • Header/body/caption prioritization

• Typed Pipeline Execution

  • All steps validated via explicit input/output types
  • Registry-driven execution model
  • No implicit coupling between pipeline stages

---

🔹 Architecture

The pipeline is built around a step registry + type-safe execution chain:

• Each step declares:

  • input type
  • output type

• The pipeline validates compatibility before execution

• Execution is explicit, deterministic, and observable

Example chain:

["preprocess", "detect_layout", "ocr_regions"]

Each step is independently replaceable and composable.

---

🔹 Key Design Decisions

Typed Data Flow

All intermediate results are structured dataclasses:

• PageImage
• PreprocessedImage
• LayoutDetection
• OCRResult

No ad-hoc dictionaries — ensures:

• traceability
• consistency
• debuggability

---

Layout-First OCR

OCR is applied after structure is understood, not before.

This prevents:

• column interleaving
• incorrect reading order
• misclassification of content

---

Fallback Over Failure

If region detection fails:

• system falls back to column-level OCR
• ensures output is still usable

---

Determinism Over Heuristics

• explicit thresholds (config-driven)
• no hidden behavior
• reproducible results across runs

---

🔹 Why This Exists

Traditional OCR pipelines:

• ignore layout
• operate on full pages
• produce inconsistent reading order
• fail silently on complex documents

This system:

• understands document structure
• isolates regions before OCR
• enforces reading order
• produces structured outputs suitable for downstream systems

---

🔹 Example Use Cases

• PDF → structured text extraction
• research document ingestion pipelines
• financial filings parsing
• multi-column article extraction
• preprocessing for NLP / LLM pipelines
• search indexing and document analysis

---

🔹 Integration Context

This module is designed to plug into:

• document ingestion systems
• OCR + NLP pipelines (e.g. abstract_hugpy)
• search and indexing systems
• large-scale document processing workflows

---

🔹 Design Philosophy

• Structure before extraction
• Determinism over convenience
• Typed pipelines over implicit flows
• Fallback over failure

---