Abstract
TabStruct-Net is a novel end-to-end deep learning framework designed to recognize table structures by integrating top-down cell detection with bottom-up graph-based association. By introducing a unique "alignment loss" and utilizing LSTMs to capture long-range visual features, the model achieves state-of-the-art results on several benchmarks without relying on PDF meta-features or OCR.
Introduction
Tables are essential components of documents, yet they present significant challenges for automated reading systems. While table detection—identifying a table's location on a page—is largely considered a solved problem, understanding internal structure remains difficult. Specifically, determining which cells belong to which rows and columns, and identifying spanning cells, is a major hurdle.
Existing methods often depend on "meta-features" like hidden text information in digital PDFs or high-quality OCR. These approaches frequently fail with scanned documents, handwritten tables, or complex layouts lacking ruling lines. TabStruct-Net addresses Physical Table Structure Recognition, converting raw images into machine-readable formats (e.g., XML) that capture both bounding boxes and logical row/column relationships.
Methodology
The TabStruct-Net framework operates through a dual-phase approach that combines object detection principles with graph-based relational modeling.
1. Top-Down: Cell Detection
The first phase utilizes a modified Mask R-CNN to localize individual cells. To accommodate the unique geometric constraints of tables, the authors introduced:
- Dilated Convolutions: These capture long-range dependencies, allowing the network to relate cells across the entire width or height of a table.
- Alignment Loss: This novel loss function penalizes the network when cells intended for the same row or column are not physically aligned in their predicted bounding boxes.
2. Bottom-Up: Structure Recognition
Once cells are detected, they are modeled as nodes in a graph to establish structural relationships.
- Visual Feature Extraction: Instead of relying solely on cell centroids, the model employs LSTMs to scan horizontal and vertical center lines, capturing richer visual context for each cell.
- Dynamic Graph CNN (DGCNN): This module predicts interactions between cells, determining if two detected cells share a row or column.
Datasets and Evaluation
The model was evaluated across diverse benchmark datasets, including both digital and scanned sources:
- ICDAR-2013 & ICDAR-2019 (cTDaR): Standard competition benchmarks.
- UNLV: Scanned document images with complex layouts.
- SciTSR & SciTSR-COMP: Large-scale scientific paper datasets.
- TableBank & PubTabNet: Massive datasets for logical structure recognition.
Key Insights
- Alignment as a Regularizer: Treating table structure as a geometric constraint forces the network to minimize coordinate variance between aligned cells, effectively teaching it the "concept" of a table grid.
- Joint Training: Training detection and association modules end-to-end improved performance by approximately 1.5% F1-score compared to disjoint training.
- Beyond Centroids: Using LSTMs to traverse the length and width of cells allows the model to better handle multi-line content and spanning cells.
Results
TabStruct-Net demonstrated significant performance gains over existing baselines:
- ICDAR-2013 (Setup-A): Using only image input, the model achieved an F1-score of 0.906, a 27.5% improvement over the previous DeepDeSRT model.
- ICDAR-2013 (Setup-B): With ground-truth cell boxes, the model reached an F1-score of 0.981.
- Robustness: The system maintained an F1-score of 0.791 even at a strict 0.7 IoU threshold, indicating high precision in cell localization.
Conclusion and Future Work
While highly effective, TabStruct-Net faces challenges with large empty spaces. Sparse tables with significant gaps can lead to localization failures in "empty" regions, breaking the association logic. Future research may focus on these sparse layouts and integrating semantic text understanding to further refine structural predictions.
Summary updated