PP-StructureV2 Table Recognition Pipeline

Relevant source files

• ppstructure/README.md
• ppstructure/table/README.md
• ppstructure/table/README_ch.md

Deprecation Notice: This document covers the PP-StructureV2 table recognition pipeline, located under ppstructure/. This directory and its associated models are no longer actively maintained and will be removed in a future release. For current table recognition capabilities, see the PP-StructureV3 document parsing system (2.3) and the specialized recognition modules overview (2.7).

---

Overview

This page documents the table recognition sub-pipeline inside the ppstructure/table/ directory. It covers:

• The three-model pipeline architecture (detection → recognition → structure)
• The SLANet table structure model
• Scripts for inference, evaluation, and label conversion
• Benchmark performance on PubTabNet

For the broader PP-StructureV2 system (layout analysis, KIE, etc.), see 8.1.

---

Pipeline Architecture

The table recognition pipeline combines three independent models that run sequentially. No single end-to-end model handles the full task; instead, OCR and table structure prediction results are merged programmatically.

Table Recognition Pipeline — Model Flow

Step-by-step description:

1. The DB model detects bounding boxes for each single-line text region in the image.
2. Detected text regions are cropped and passed to the CRNN model for text recognition.
3. SLANet independently processes the full image and outputs table structure tokens (HTML tags) plus bounding boxes for each cell.
4. Text recognition results from step 2 are matched to cell bounding boxes from step 3 using coordinate overlap.
5. Cell text and table structure tokens are combined to produce a complete HTML table string, which is then converted to Excel.

Sources: ppstructure/table/README.md15-28

---

Models

Three distinct model types are required. Each is loaded from a separate inference model directory.

Role	Model Type	Example Model Name	Dictionary File
Text Detection	DB	PP-OCRv3_mobile_det_infer	—
Text Recognition	CRNN	ch_PP-OCRv3_rec_infer	ppocr_keys_v1.txt
Table Structure	SLANet	ch_ppstructure_mobile_v2.0_SLANet_infer	table_structure_dict_ch.txt

English variants use en_ppocr_mobile_v2.0_table_det_infer, en_ppocr_mobile_v2.0_table_rec_infer, and en_ppstructure_mobile_v2.0_SLANet_infer with table_structure_dict.txt.

The table structure dictionary files live at:

• ppocr/utils/dict/table_structure_dict_ch.txt (Chinese/SLANet)
• ppocr/utils/dict/table_structure_dict.txt (English/SLANet or TableRec-RARE)
• ppocr/utils/dict/table_dict.txt (used during TEDS evaluation)

Sources: ppstructure/table/README.md54-88

---

Script Reference

Code entity map — scripts and their roles

Sources: ppstructure/table/README.md52-155

predict_table.py

Main inference script. Accepts all three model directories as arguments and produces per-image Excel files plus an HTML visualization.

Argument	Description
--det_model_dir	Path to the text detection (DB) inference model
--rec_model_dir	Path to the text recognition (CRNN) inference model
--table_model_dir	Path to the SLANet table structure inference model
--rec_char_dict_path	Character dictionary for the recognition model
--table_char_dict_path	Structure token dictionary for SLANet
--image_dir	Input image or directory
--output	Output directory for Excel and HTML files
--merge_no_span_structure	Set to False when using TableRec-RARE model

Sources: ppstructure/table/README.md60-88

eval_table.py

Evaluates the full pipeline against a ground-truth file using the TEDS metric. Requires all three model directories and a prepared GT file.

Additional arguments beyond those shared with predict_table.py:

Argument	Description
--det_limit_side_len	Detection image side limit (e.g., 736)
--det_limit_type	Limit type: min or max
--rec_image_shape	Recognition input shape, e.g., 3,32,320
--gt_path	Path to the GT label file

Sources: ppstructure/table/README.md98-155

convert_label2html.py

Utility that converts an annotation label file into the HTML-based GT format required by eval_table.py.

python3 ppstructure/table/convert_label2html.py \
    --ori_gt_path /path/to/your_label_file \
    --save_path /path/to/save_file

The GT file format is one entry per line: <filename>\t<html_table_string>.

Sources: ppstructure/table/README.md100-112

---

Benchmark Performance

Evaluated on the PubTabNet validation dataset. All speeds measured on CPU with MKL enabled.

Method	Acc	TEDS	Speed (CPU)
EDD	—	88.30%	—
TableRec-RARE	71.73%	93.88%	779 ms/image
SLANet	76.31%	95.89%	766 ms/image

• Acc: Exact match accuracy per image; any incorrect token counts as a full error.
• TEDS (Tree-Edit-Distance-based Similarity): Measures similarity of the predicted HTML tree to the ground-truth tree, covering both structure and text content.

Sources: ppstructure/table/README.md31-43

---

Evaluation Workflow

Sources: ppstructure/table/README.md98-155

---

Migration to PP-StructureV3

The ppstructure/ directory is pending removal. Key replacements:

PP-StructureV2 Capability	Recommended Replacement
Table recognition via SLANet	Table recognition module in PP-StructureV3 (2.3)
KIE (Key Information Extraction)	PP-ChatOCRv4 (2.4)
ppstructure/table/predict_table.py	PP-StructureV3 integrated wheel package

Sources: ppstructure/README.md1-16