Skip to main content

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

Metadata

Release files for abstract-ocr 0.0.1.71

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for abstract-ocr 0.0.1.71
File Size Uploaded Provenance Actions
abstract_ocr-0.0.1.71.tar.gz 52.3 kB Without provenance Details for abstract_ocr-0.0.1.71.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for abstract-ocr 0.0.1.71
File Interpreter ABI Platform Actions
abstract_ocr-0.0.1.71-py3-none-any.whl Python 3 none any Details for abstract_ocr-0.0.1.71-py3-none-any.whl

Total release size: 125.2 kB

Release files / abstract_ocr-0.0.1.71.tar.gz

Download URL abstract_ocr-0.0.1.71.tar.gz
Size 52.3 kB
Tags Source
SHA-256 checksum
How to use checksums
e8ca1d985989e9402bd32561bd6a87a55107af52f26cca010547882781039958
BLAKE2b-256 checksum
How to use checksums
01804aac0dbaf813025d9cf02983a8d54ee8b8018c7636ee3e0bc9c3964ab47c
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.12.3

Release files / abstract_ocr-0.0.1.71-py3-none-any.whl

Download URL abstract_ocr-0.0.1.71-py3-none-any.whl
Size 72.9 kB
Tags Python 3
SHA-256 checksum
How to use checksums
fd657f9a30ba2928ecb1b5d1fea7af6f27f89a984dfa18d48820af7c80468d2c
BLAKE2b-256 checksum
How to use checksums
b5b520b9d08f9666366b158bebcc4a9a80e0111b934ff1617840f16c1208cb64
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.12.3

Security for abstract-ocr 0.0.1.71

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release

0.0.1.71 This release

2 release files No provenance

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page