Open-source tatr_tab_struct_v2 model - Efficiently accomplish table structure recognition tasks, available for free!

Tatr Tab Struct V2

Developed by deepdoctection

DETR architecture model trained on PubTables1M and FinTabNet datasets, specialized for table structure recognition tasks

Text Recognition

Transformers

#Table Structure Recognition #Cross-cell Detection #Financial Document Processing

Downloads 99

Release Time : 9/4/2023

Model Overview

This model adopts the Transformer architecture, capable of identifying rows, columns, headers, and cross-cell structures in tables, suitable for document digitization scenarios

Model Features

Cross-cell Recognition

Accurately identifies merged cells and cross-row/column structures in tables

Multi-element Detection

Simultaneously detects various layout elements such as tables, rows, columns, and headers

Optimized Edge Processing

Recommended to use a 5-pixel padding margin for optimal recognition results

Model Capabilities

Table Region Detection

Row/Column Structure Recognition

Header Classification

Merged Cell Detection

Use Cases

Document Digitization

Financial Statement Parsing

Automatically identifies structured data in complex financial statements

Accurately extracts row/column relationships and cross-cell data

Academic Literature Processing

Extracts table content from research papers

Preserves the hierarchical relationships of original tables

🚀 Transformers

Microsoft Table Transformer for table structure recognition, trained on Pubtables and Fintabnet.

🚀 Quick Start

If you don't have the deepdoctection profile of the model, you need to add the following code:

Basic Usage

import deepdoctection as dd

dd.ModelCatalog.register("deepdoctection/tatr_tab_struct_v2/pytorch_model.bin", dd.ModelProfile(
    name="deepdoctection/tatr_tab_struct_v2/pytorch_model.bin",
    description="Table Transformer (DETR) model trained on PubTables1M. It was introduced in the paper "
                "Aligning benchmark datasets for table structure recognition by Smock et "
                "al. This model is devoted to table structure recognition and assumes to receive a slightly cropped"
                "table as input. It will predict rows, column and spanning cells. Use a padding of around 5 pixels",
    size=[115511753],
    tp_model=False,
    config="deepdoctection/tatr_tab_struct_v2/config.json",
    preprocessor_config="deepdoctection/tatr_tab_struct_v2/preprocessor_config.json",
    hf_repo_id="deepdoctection/tatr_tab_struct_v2",
    hf_model_name="pytorch_model.bin",
    hf_config_file=["config.json", "preprocessor_config.json"],
    categories={
        "1": dd.LayoutType.table,
        "2": dd.LayoutType.column,
        "3": dd.LayoutType.row,
        "4": dd.CellType.column_header,
        "5": dd.CellType.projected_row_header,
        "6": dd.CellType.spanning,
    },
    dl_library="PT",
    model_wrapper="HFDetrDerivedDetector",
))

When running the model within the deepdoctection analyzer, you should adjust the segmentation parameters to get better predictions.

Advanced Usage

import deepdoctection as dd

analyzer = dd.get_dd_analyzer(reset_config_file=True, config_overwrite=["PT.ITEM.WEIGHTS=deepdoctection/tatr_tab_struct_v2/pytorch_model.bin",
                                                                        "PT.ITEM.FILTER=['table']",
                                                                        "PT.ITEM.PAD.TOP=5",
                                                                        "PT.ITEM.PAD.RIGHT=5",
                                                                        "PT.ITEM.PAD.BOTTOM=5",
                                                                        "PT.ITEM.PAD.LEFT=5",
                                                                        "SEGMENTATION.THRESHOLD_ROWS=0.9",
                                                                        "SEGMENTATION.THRESHOLD_COLS=0.9",
                                                                        "SEGMENTATION.REMOVE_IOU_THRESHOLD_ROWS=0.3",
                                                                        "SEGMENTATION.REMOVE_IOU_THRESHOLD_COLS=0.3",
                                                                        "WORD_MATCHING.MAX_PARENT_ONLY=True"])

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご