Table Recognition

Relevant source files

• mineru/model/table/cls/paddle_table_cls.py
• mineru/model/table/rec/onnxruntime_provider.py
• mineru/model/table/rec/slanet_plus/__init__.py
• mineru/model/table/rec/slanet_plus/main.py
• mineru/model/table/rec/slanet_plus/matcher.py
• mineru/model/table/rec/slanet_plus/matcher_utils.py
• mineru/model/table/rec/slanet_plus/table_structure.py
• mineru/model/table/rec/slanet_plus/table_structure_utils.py
• mineru/model/table/rec/unet_table/__init__.py
• mineru/model/table/rec/unet_table/main.py
• mineru/model/table/rec/unet_table/table_recover.py
• mineru/model/table/rec/unet_table/table_structure_unet.py
• mineru/model/table/rec/unet_table/utils.py
• mineru/model/table/rec/unet_table/utils_table_line_rec.py
• mineru/model/table/rec/unet_table/utils_table_recover.py

The table recognition system in MinerU is a sophisticated dual-path pipeline designed to handle both "Wired" (tables with explicit grid lines) and "Wireless" (tables relying on whitespace alignment) structures. It incorporates automated orientation correction, specialized classification, and a custom grid reconstruction algorithm to produce high-fidelity HTML and Markdown representations of tabular data.

System Architecture

The table recognition logic is orchestrated within the MineruPipelineModel and BatchAnalyze classes. When a table region is detected by the layout model, the system extracts the table image and routes it through a series of specialized models managed by the AtomModelSingleton.

Table Recognition Data Flow

The following diagram illustrates the lifecycle of a table from a raw image to a reconstructed HTML structure, mapping high-level steps to their corresponding code entities.

Diagram: Table Processing Pipeline

Sources: mineru/model/table/rec/unet_table/main.py68-129 mineru/model/table/rec/slanet_plus/main.py165-187 mineru/model/table/cls/paddle_table_cls.py134-150

---

Orientation & Classification

Before recognition begins, classification steps ensure the table image is in the optimal state for the recognition models.

1. Orientation Correction

The orientation logic ensures tables are upright before structural analysis.

• Model: MineruTableOrientationClsModel mineru/model/table/cls/mineru_table_ori_cls.py25-27
• Logic: It uses a dual-gated approach. First, it identifies "rotation candidates" by detecting vertical (high-aspect-ratio) text boxes mineru/model/table/cls/mineru_table_ori_cls.py68-90 If a candidate is found, it performs OCR recognition on three rotations (0°, 90°, 270°) and selects the orientation with the highest confidence score mineru/model/table/cls/mineru_table_ori_cls.py166-172
• Parameters: It samples up to 18 boxes for scoring mineru/model/table/cls/mineru_table_ori_cls.py18 and uses a priority threshold of 0.9 for the 0° orientation to avoid unnecessary rotations mineru/model/table/cls/mineru_table_ori_cls.py20

2. Wired vs. Wireless Classification

The PaddleTableClsModel mineru/model/table/cls/paddle_table_cls.py16-26 determines which recognition path to take by classifying the table structure.

• Wired Tables: Characterized by visible cell borders. Handled by the UNet-based TSRUnet.
• Wireless Tables: Characterized by open structures relying on alignment. Handled by SlanetPlus or RapidTable.
• Implementation: Uses an ONNX inference session (onnxruntime.InferenceSession) to predict the class and confidence score mineru/model/table/cls/paddle_table_cls.py18-20 The model accepts a batch of images and updates the table_res dictionary with cls_label and cls_score mineru/model/table/cls/paddle_table_cls.py134-150
• Preprocessing: Images are resized to 224x224 and normalized using mean [0.485, 0.456, 0.406] and std [0.229, 0.224, 0.225] mineru/model/table/cls/paddle_table_cls.py21-25

Sources: mineru/model/table/cls/mineru_table_ori_cls.py25-192 mineru/model/table/cls/paddle_table_cls.py16-76 mineru/model/table/cls/paddle_table_cls.py134-150

---

Wired Table Recognition (TSRUnet)

The wired path uses a UNet-based architecture to segment table lines and a geometric recovery algorithm to reconstruct the grid.

TSRUnet Model

The TSRUnet class mineru/model/table/rec/unet_table/table_structure_unet.py25-38 performs:

1. Inference: Predicts a segmentation mask where pixel values represent horizontal (1) or vertical (2) lines mineru/model/table/rec/unet_table/table_structure_unet.py79-82
2. Line Extraction: Uses morphological operations (cv2.morphologyEx) and get_table_line mineru/model/table/rec/unet_table/utils_table_line_rec.py129-146 to extract discrete line segments from the mask mineru/model/table/rec/unet_table/table_structure_unet.py117-128
3. Refinement: Adjusts and extends lines to ensure intersections are closed using final_adjust_lines and adjust_lines mineru/model/table/rec/unet_table/table_structure_unet.py131-137
4. Rotation Fix: Calculates the rotation angle of the table lines via cal_rotate_angle mineru/model/table/rec/unet_table/table_structure_unet.py164-177 and optionally performs rotate_image to normalize the grid mineru/model/table/rec/unet_table/table_structure_unet.py140-146

TableRecover & Grid Reconstruction

The WiredTableRecognition class mineru/model/table/rec/unet_table/main.py61-66 orchestrates the full recovery:

• TableRecover: Transforms predicted polygons into a logical grid with logic_points mineru/model/table/rec/unet_table/main.py89-91
• Cell Matching: Maps OCR results to predicted polygons using match_ocr_cell mineru/model/table/rec/unet_table/main.py107
• Blank Filling: If a detected cell has no OCR content, the system crops the image and performs targeted OCR via fill_blank_rec mineru/model/table/rec/unet_table/main.py109 This includes a noise filter _should_drop_blank_cell_rec_result to drop low-confidence or irrelevant characters mineru/model/table/rec/unet_table/main.py174-184
• HTML Plotting: Final HTML is generated by plot_html_table mineru/model/table/rec/unet_table/main.py121

Sources: mineru/model/table/rec/unet_table/main.py46-129 mineru/model/table/rec/unet_table/table_structure_unet.py25-150 mineru/model/table/rec/unet_table/utils_table_line_rec.py129-175 mineru/model/table/rec/unet_table/utils_table_recover.py122-158

---

Wireless Table Recognition (SlanetPlus)

Wireless tables are processed using the PaddleTableModel wrapper around the SlanetPlus architecture.

TableStructurer

The TableStructurer mineru/model/table/rec/slanet_plus/table_structure.py28-40 uses an ONNX model to predict both the table structure (HTML tags) and the cell bounding boxes.

• Output: A sequence of structure tokens (e.g., <tr>, <td>, </td>) and a list of cell coordinates mineru/model/table/rec/slanet_plus/table_structure.py53-68
• Batch Processing: batch_process enables high-throughput inference across multiple table images using BatchTablePreprocess mineru/model/table/rec/slanet_plus/table_structure.py70-113

TableMatch & HTML Reconstruction

The PaddleTable class aligns OCR text with the predicted structure:

1. Normalization: Standardizes cell bboxes and restores scale via adapt_slanet_plus mineru/model/table/rec/slanet_plus/main.py137-145
2. Vectorized Matching: Efficiently calculates mapping between OCR results and predicted structure using match_result mineru/model/table/rec/slanet_plus/matcher.py132-159 It uses IoU and distance scores to pair OCR boxes with cell boxes mineru/model/table/rec/slanet_plus/matcher.py73-102
3. Logic Point Decoding: Extracts logical coordinates (row/column spans) using decode_logic_points mineru/model/table/rec/slanet_plus/main.py68
4. HTML Generation: The table_matcher generates the final HTML string by combining predicted structures, cell bboxes, and OCR results mineru/model/table/rec/slanet_plus/main.py62

Sources: mineru/model/table/rec/slanet_plus/main.py37-71 mineru/model/table/rec/slanet_plus/table_structure.py28-113 mineru/model/table/rec/slanet_plus/matcher.py120-159

---

Model Arbitration Logic

The system uses the AtomModelSingleton to manage the lifecycle of these models. The specific recognition model is selected based on the output of the classification step.

Diagram: Code Entity Mapping

System Concept	Implementation Class	Primary File
Model Manager	AtomModelSingleton	mineru/backend/pipeline/model_init.py
Wired Recognition	WiredTableRecognition	mineru/model/table/rec/unet_table/main.py61-66
Wireless Recognition	PaddleTableModel	mineru/model/table/rec/slanet_plus/main.py153-163
Structure Recovery	TableRecover	mineru/model/table/rec/unet_table/main.py65
Table Classifier	PaddleTableClsModel	mineru/model/table/cls/paddle_table_cls.py16-26
Table Structurer	TableStructurer	mineru/model/table/rec/slanet_plus/table_structure.py28-40
Inference Session	OrtInferSession	mineru/model/table/rec/unet_table/utils.py26-41

Table Merging and Continuation

For tables spanning multiple pages, MinerU implements specialized merging logic in mineru/utils/table_merge.py.

• Row Scanning: _scan_rows calculates effective column metrics and preserved rowspans across boundaries mineru/utils/table_merge.py78-146
• Continuation Detection: is_table_continuation_text and _is_continuation_caption identify "Continued Table" markers in captions mineru/utils/table_merge.py11 mineru/utils/table_merge.py203-205
• State Management: TableMergeState caches header signatures and row metrics to determine if a subsequent table is a logical continuation of the previous one mineru/utils/table_merge.py54-67

Sources: mineru/utils/table_merge.py11-205 mineru/model/table/rec/unet_table/main.py46-51 mineru/model/table/rec/slanet_plus/main.py153-163 mineru/model/table/rec/unet_table/utils.py26-41 mineru/model/table/rec/slanet_plus/table_structure.py28-40 mineru/model/table/cls/paddle_table_cls.py16-76