microsoft/table-transformer
Table Transformer
Table Transformer is a deep learning framework designed for document layout analysis and the automated extraction of tabular data from unstructured images and documents. It utilizes a transformer-based architecture to perform object detection, identifying table boundaries and internal grid structures to convert visual information into machine-readable formats.
The project distinguishes itself by employing a global optimization strategy for bipartite matching, which eliminates the need for traditional non-maximum suppression during the detection process. By leveraging multi-scale feature extraction and self-attention mechanisms, the model maps raw visual inputs directly to structured coordinate data, allowing for the precise localization of rows, columns, and individual cells within complex layouts.
The library supports the full lifecycle of document analysis, including the training and fine-tuning of models on custom datasets to accommodate specialized document domains. It provides integrated tools for evaluating extraction performance, utilizing standard detection metrics and grid similarity scores to assess the accuracy of structure recognition against ground truth data.
Features
- Document Table Extractors - Extracts structured tabular data from unstructured PDF and image files for analysis and storage.
- Transformer-Based Architectures - Implements transformer-based object detection to predict bounding boxes and class labels for table components.
- Table Structure Detections - Locates table structures within unstructured images by identifying spatial boundaries of tabular content.
- Visual Table Extraction - Converts detected table regions into structured formats by mapping rows, columns, and cell boundaries.
- Bounding Box Detection - Identifies and localizes table components within visual media using neural network-based bounding box detection.
- Object Detection - Provides a neural network architecture for localizing and classifying table regions using precise bounding boxes.
- Vision Detection Model Training - Supports training and fine-tuning of specialized detection models on domain-specific datasets by adjusting hyperparameters.
- Document Analysis Models - Offers a framework for training and evaluating specialized models on document layouts to improve spatial recognition.
- Encoder-Decoder Architectures - Utilizes an encoder-decoder transformer architecture to map visual image patches into structured coordinate data.
- Multi-Scale Feature Pyramids - Extracts multi-scale visual representations to detect table elements of varying sizes and complex spatial layouts.
- Hungarian Matching Losses - Employs a global bipartite matching strategy to assign predictions to ground truth objects without non-maximum suppression.
- Machine Learning Training - Enables fine-tuning of deep learning models on custom datasets to improve detection accuracy for specialized tabular formats.
- Document Layout Analysis - Provides tools for document layout analysis to identify and locate tables within unstructured pages.
- Detection Accuracy Evaluators - Evaluates extraction performance using standard detection metrics and specialized grid similarity scores against ground truth data.
- Supervised Fine-Tuning - Supports supervised fine-tuning of pre-trained models on labeled datasets of structured table layouts.
- Table Grid Mappings - Maps detected cell coordinates into a logical matrix format by analyzing spatial relationships between table components.
