Skip to content

Detectors

Available layout detection models and their characteristics.

Overview

Layout detectors identify document regions (text, tables, figures, etc.) in page images. The pipeline supports multiple detector backends with different trade-offs between speed and accuracy.

Detector Comparison

Detector Source Technology Speed Quality Block Types
doclayout-yolo This project YOLO v8 ⚡⚡⚡ ⭐⭐⭐ 7
mineru-doclayout-yolo MinerU YOLO v8 ⚡⚡ ⭐⭐⭐ 10
paddleocr-doclayout-v2 PaddleOCR PP-DocLayoutV2 ⚡⚡ ⭐⭐⭐⭐ 25
mineru-vlm MinerU VLM-based ⚡ ⭐⭐⭐⭐ 25+

DocLayout-YOLO

Speed: Fast | Quality: Good | GPU Memory: ~2GB

This project's optimized DocLayout-YOLO detector based on Ultralytics YOLO v8.

Block Types

Type Description
title Document titles and headings
plain text / text Regular paragraph text
figure Images and diagrams
table Tables
equation Mathematical formulas
list Bullet/numbered lists

Usage

from pipeline.layout.detection import create_detector

detector = create_detector(
    "doclayout-yolo",
    model_path="path/to/model.pt",  # Optional custom model
    confidence_threshold=0.5,
    auto_batch_size=True,  # Auto-optimize batch size
)

blocks = detector.detect(image)

CLI

python main.py --input doc.pdf --detector doclayout-yolo --confidence 0.5

PaddleOCR PP-DocLayoutV2

Speed: Medium | Quality: Very Good | GPU Memory: ~4GB

PaddleOCR's PP-DocLayoutV2 detector with 25 category types and built-in pointer network for reading order.

Block Types (25 categories)

Category Types
Titles doc_title, paragraph_title
Text text, sidebar_text, abstract, contents
References reference, reference_content, footnote
Page Elements header, footer, header_image, footer_image, page_number
Formulas formula, formula_number, algorithm
Tables table, table_title
Images image, chart, chart_title, seal

Usage

detector = create_detector(
    "paddleocr-doclayout-v2",
    confidence_threshold=0.5,
)

CLI

python main.py --input doc.pdf --detector paddleocr-doclayout-v2

Requirements

  • PaddleOCR v3.3.0+ (in external/PaddleOCR/)
  • PaddlePaddle with CUDA support (recommended)

MinerU DocLayout-YOLO

Speed: Fast | Quality: Good | GPU Memory: ~2GB

MinerU's implementation of DocLayout-YOLO with additional block types.

Block Types

Type Description
title Document titles
plain text Regular text
abandon Discarded regions
figure Images and diagrams
figure_caption Figure captions
table Tables
table_caption Table captions
table_footnote Table footnotes
isolate_formula Standalone formulas
formula_caption Formula captions

Usage

detector = create_detector("mineru-doclayout-yolo")

CLI

python main.py --input doc.pdf --detector mineru-doclayout-yolo

Requirements

  • MinerU 2.5+ (in external/MinerU/)

MinerU VLM

Speed: Slow | Quality: Excellent | GPU Memory: ~8GB

Vision Language Model-based detection using MinerU's VLM pipeline.

Block Types

Full MinerU 2.5 hierarchy with 25+ block types including:

  • Content: text, title, list
  • Images: image, image_body, image_caption, image_footnote
  • Tables: table, table_body, table_caption, table_footnote
  • Equations: interline_equation, inline_equation
  • Code: code, code_body, code_caption, algorithm
  • Page elements: header, footer, page_number, page_footnote
  • References: ref_text, phonetic, aside_text, index

Usage

detector = create_detector(
    "mineru-vlm",
    model_name="opendatalab/MinerU2.5-2509-1.2B",
    backend="hf",  # or "vllm" for faster inference
)

CLI

python main.py --input doc.pdf --detector mineru-vlm --detector-backend vllm

Requirements

  • MinerU 2.5+ (in external/MinerU/)
  • Large GPU (8GB+ VRAM recommended)

Backend Selection

Detectors can use different inference backends:

Backend Description Supported By
pytorch Native PyTorch doclayout-yolo, mineru-doclayout-yolo
paddle PaddlePaddle paddleocr-doclayout-v2
hf HuggingFace Transformers mineru-vlm
vllm vLLM high-throughput mineru-vlm
pt-ray PyTorch + Ray multi-GPU doclayout-yolo 
# Explicit backend selection
python main.py --input doc.pdf --detector mineru-vlm --detector-backend vllm

Choosing a Detector

Use Case Detector Reason
Fast processing doclayout-yolo Fastest, good quality
High quality paddleocr-doclayout-v2 25 types, built-in ordering
Complex documents mineru-vlm Best understanding, slowest
MinerU integration mineru-doclayout-yolo MinerU ecosystem

Decision Matrix

graph TD
    A[Need fast processing?] -->|Yes| B[doclayout-yolo]
    A -->|No| C[Need detailed types?]
    C -->|Yes| D[paddleocr-doclayout-v2]
    C -->|No| E[Complex layouts?]
    E -->|Yes| F[mineru-vlm]
    E -->|No| B

Performance Benchmarks

Tested on NVIDIA RTX 3090 with 300 DPI images:

Detector Pages/min GPU Memory Init Time
doclayout-yolo ~60 2 GB 2s
paddleocr-doclayout-v2 ~40 4 GB 5s
mineru-doclayout-yolo ~50 2 GB 3s
mineru-vlm ~10 8 GB 30s

See Also