Model Subsystems

Relevant source files

• mineru/backend/pipeline/batch_analyze.py
• mineru/backend/pipeline/model_init.py
• mineru/backend/pipeline/model_list.py
• mineru/model/table/cls/mineru_table_ori_cls.py

MinerU utilizes a collection of specialized machine learning models to perform document decomposition. These models are managed as atomic units and orchestrated by the pipeline to transform raw page images into structured data. This page provides a high-level overview of these subsystems and their roles within the MinerU architecture.

The lifecycle of these models is managed by the AtomModelSingleton mineru/backend/pipeline/model_init.py148-157 which ensures that heavy model weights are loaded into memory only once and shared across processing tasks using a thread-safe implementation mineru/backend/pipeline/model_init.py184-187

Subsystem Architecture Overview

The following diagram illustrates how the different model subsystems interact during the processing of a single document page within the BatchAnalyze orchestration layer mineru/backend/pipeline/batch_analyze.py52-81

Model Interaction Flow

Sources: mineru/backend/pipeline/batch_analyze.py52-81 mineru/backend/pipeline/model_init.py189-220

Model Entity Mapping

This diagram maps system functions to their specific implementation classes and file paths within the codebase, as resolved by the AtomModelSingleton mineru/backend/pipeline/model_init.py159-187

Sources: mineru/backend/pipeline/model_init.py159-220 mineru/backend/pipeline/model_list.py2-10 mineru/model/table/cls/mineru_table_ori_cls.py25-27

Layout Detection & Reading Order

The layout detection subsystem identifies functional regions such as text blocks, titles, figures, tables, and formulas.

• Primary Model: PPDocLayoutV2LayoutModel, which categorizes regions into functional labels like doc_title, table, and display_formula.
• Initialization: Managed via pp_doclayout_v2_model_init mineru/backend/pipeline/model_init.py126-130
• Post-processing: The system removes empty OCR text blocks during analysis via _prune_empty_ocr_text_blocks mineru/backend/pipeline/batch_analyze.py166-180
• Concurrency: Layout inference is protected by PIPELINE_LAYOUT_INFERENCE_LOCK mineru/backend/pipeline/model_init.py24 when enabled via environment variables mineru/backend/pipeline/model_init.py28-30

For details, see Layout Detection & Reading Order.

Table Recognition

MinerU employs a dual-path recognition strategy to handle the structural diversity of tables.

• Orientation Correction: Detects if table images are rotated and corrects them using MineruTableOrientationClsModel mineru/backend/pipeline/model_init.py72-82 It uses OCR scores to determine the final rotation angle (0, 90, or 270 degrees) mineru/model/table/cls/mineru_table_ori_cls.py22
• Classification: PaddleTableClsModel determines if a table is WiredTable (has visible grid lines) or WirelessTable.
• Wired Tables: Processed via UnetTableModel mineru/backend/pipeline/model_init.py98 which utilizes a UNet-based architecture.
• Wireless Tables: Processed via PaddleTableModel mineru/backend/pipeline/model_init.py111 utilizing the SlanetPlus architecture.

For details, see Table Recognition.

OCR Engine

The OCR subsystem handles text extraction from non-digitized PDF regions and image-based content.

• Engine: PytorchPaddleOCR, initialized via ocr_model_init mineru/backend/pipeline/model_init.py133-145
• Optimization: Supports batch detection and recognition controlled by constants like OCR_DET_BASE_BATCH_SIZE mineru/backend/pipeline/batch_analyze.py40
• Language Support: Configured during initialization with support for normalized language codes mineru/backend/pipeline/model_init.py139
• Masking: Can mask inline formulas to improve OCR accuracy for surrounding text via mask_formula_regions_for_ocr_det mineru/backend/pipeline/batch_analyze.py88

For details, see OCR Engine.

Formula Recognition (MFD/MFR)

Mathematical content is handled by a specialized two-stage pipeline: Mathematical Formula Detection (MFD) and Mathematical Formula Recognition (MFR).

• Recognition (MFR): MinerU supports two primary recognition models initialized via mfr_model_init mineru/backend/pipeline/model_init.py115-123:
  • UnimernetModel: The default model for general formula recognition mineru/backend/pipeline/model_init.py117
  • FormulaRecognizer: PP-FormulaNet-Plus-M, used when Chinese formula support is enabled mineru/backend/pipeline/model_init.py119
• Batching: Inference can be batched with a base size of 16 mineru/backend/pipeline/batch_analyze.py39

For details, see Formula Recognition (MFD/MFR).

Model Management & Initialization

Models are initialized through a centralized factory pattern in atom_model_init mineru/backend/pipeline/model_init.py189-220 to manage device placement and lifecycle.

Model Category	Class Name	Initialization Function
Layout	PPDocLayoutV2LayoutModel	pp_doclayout_v2_model_init mineru/backend/pipeline/model_init.py126
MFR	UnimernetModel / FormulaRecognizer	mfr_model_init mineru/backend/pipeline/model_init.py115
OCR	PytorchPaddleOCR	ocr_model_init mineru/backend/pipeline/model_init.py133
Table Cls	PaddleTableClsModel	table_cls_model_init mineru/backend/pipeline/model_init.py85
Wired Table	UnetTableModel	wired_table_model_init mineru/backend/pipeline/model_init.py89
Wireless Table	PaddleTableModel	wireless_table_model_init mineru/backend/pipeline/model_init.py102
Orientation	MineruTableOrientationClsModel	table_orientation_cls_model_init mineru/backend/pipeline/model_init.py72

Sources: mineru/backend/pipeline/model_init.py1-220 mineru/backend/pipeline/batch_analyze.py38-50 mineru/backend/pipeline/model_list.py1-11 mineru/model/table/cls/mineru_table_ori_cls.py13-22