Source-linked AI summary
TNCR: Table Net Detection and Classification Dataset
Abdelrahman Abdallah, Alexander Berendeyev, Islam Nuradin, Daniyar Nurseitov
TL;DR
Table detection and structure recognition are difficult because document tables vary in format and layout and can be confused with other page content. The paper introduces TNCR, a real-image dataset for table detection and five-class classification, and evaluates deep-learning baselines, with Cascade Mask R-CNN achieving the highest reported F1 score among the compared methods.
Problem
Document table detection is difficult because formats and layouts vary, and tables may be confused with text, figures, split columns or rows, nested tables, and embedded figures.
Source: §1 · p. 2Method
Results
0.844 F1 score: Cascade Mask R-CNN achieved this score, compared with 0.841 for Cascade R-CNN, 0.804 for Cascade RPN, 0.840 for Hybrid Task Cascade, and 0.492 for YOLO.
Source: §6 · p. 19Takeaways & Limitations
The authors conclude that deep learning for detecting and recognizing tables from images is a promising research direction and that TNCR can aid table analysis research.
Source: §6 · pp. 18–19
Abstract
from arXiv · showhide
We present TNCR, a new table dataset with varying image quality collected from free websites. The TNCR dataset can be used for table detection in scanned document images and their classification into 5 different classes. TNCR contains 9428 high-quality labeled images. In this paper, we have implemented state-of-the-art deep learning-based methods for table detection to create several strong baselines. Cascade Mask R-CNN with ResNeXt-101-64x4d Backbone Network achieves the highest performance compared to other methods with a precision of 79.7%, recall of 89.8%, and f1 score of 84.4% on the TNCR dataset. We have made TNCR open source in the hope of encouraging more deep learning approaches to table detection, classification, and structure recognition. The dataset and trained model checkpoints are available at https://github.com/abdoelsayed2016/TNCR_Dataset.
1. Introduction
Table detection and structure recognition are difficult because document formats and layouts vary, and tables can be confused with text or figures. The paper introduces TNCR and evaluates deep-learning detection and classification baselines on real table images.
1. Introduction
- Document-table detection is difficult across varied formats and layouts, especially with split rows or columns, nested tables, and embedded figures.Tables must also be distinguished from text and other figures.Source: §1 · p. 2
- The paper applies instance segmentation to detect tables at pixel level and uses the same model to classify segmented tables into five classes.The authors describe table detection and classification as the two tasks addressed with deep-learning models.Source: §1 · p. 2
- TNCR provides real images of varying quality for table detection and classification into five classes.The images are not generated from LaTeX or Word; the classes are full lined, no lines, merged cells, partial lined, and partial lined merged cells.Source: §1 · pp. 2–3
- The baselines compare Cascade R-CNN, Cascade Mask R-CNN, Cascade RPN, Hybrid Task Cascade, and YOLO across ResNet and ResNeXt backbones and training schedules.The paper also presents COCO performance metrics over IoUs from 50% to 95%.Source: §1 · p. 4
2. Related Work
The related work surveys table datasets and methods for locating tables and recovering their row-and-column structure. It covers conventional image-processing and machine-learning approaches as well as deep-learning systems.
2.1. Existing Datasets
- Existing datasets include scanned-page collections and table-detection benchmarks, while DeepFigures focuses on detection and cannot be used for table structure recognition.TableBank is described as 417K high-quality labeled tables across domains, with their original documents.Source: §2.1 · pp. 4–5
2.2. Table detection and structure detection
- Earlier structure methods infer rows, columns, cells, and junctions from lines, bounding boxes, or layout components.The cited approaches include expanding cell boxes, labeling edges, and recognizing spanning cells.Source: §2.2 · pp. 5–6
- Machine-learning approaches include hierarchical page representations, SVM feature extraction, and HMM-based table location.The cited history also describes unsupervised and supervised methods for table detection.Source: §2.2 · p. 6
- Deep-learning systems use CNN-based detection, adaptive receptive fields, or joint table and structure models.Examples include Faster R-CNN, deformable CNNs, CascadeTabNet, DeepDeSRT, TableNet, and a saliency-based fully convolutional model.Source: §2.2 · pp. 6–7
3. Table Net Detection and Classification Dataset (TNCR)
TNCR groups tables into five visual classes based on line presence and merged cells. The paper also describes augmenting underrepresented classes with tables found by parsing public FDA PDF documents.
3. Table Net Detection and Classification Dataset (TNCR)
- TNCR distinguishes five table types by line completeness and whether cells are merged.The classes are full lined, no lines, merged cells, partial lined, and partial lined merged cells.Source: §3 · p. 8
- The classes distinguish fully ruled tables, unruled tables, tables with merged cells, and tables with missing rules, including a mixed missing-rule-and-merged-cell class.A merged cell joins two or more cells without a dividing line; partial-lined tables have one or more lines missing.Source: §3 · p. 8
- To supplement three classes with too few tables for balance, the authors parsed 875,026 FDA PDF pages and used a Faster R-CNN model to identify 225,154 pages with tables.Tables from those pages were used to add examples to the underrepresented classes and repartition the dataset.Source: §3 · p. 8
4. Methodology
The methodology describes object-detection models for table detection and classification. It covers cascaded detectors, segmentation extensions, and YOLO, with methods designed to refine proposals or add segmentation and context.
4.1. Cascade R-CNN
- Cascade R-CNN raises the IoU threshold across stages and reuses refined boxes as proposals; the authors report no further improvement after three stages.Its end-to-end training helps later stages discard low-quality proposals from earlier stages.Source: §4.1 · pp. 9–10
4.2. Cascade Mask R-CNN
- Cascade Mask R-CNN adds a mask-segmentation branch parallel to box regression and classification, placing it at the first stage to limit computation.The passage notes that pixel-wise segmentation is not necessarily improved by a better-defined bounding box.Source: §4.2 · p. 11
4.3. Cascade RPN
- Cascade RPN uses adaptive convolution to align features with anchors; its first-stage dilated convolution preserves spatial order and bridges features to later stages.Source: §4.3 · pp. 11–12
4.4. Hybrid Task Cascade (HTC)
- HTC combines cascade stages and multitasking with spatial context, interleaving box regression and mask prediction and adding a semantic segmentation branch.It also feeds previous-stage mask features into the current stage.Source: §4.4 · pp. 12–13
4.5. YOLO
- YOLOv3 assigns one prior box to each ground-truth object; it ignores some overlapping non-best priors and applies a 0.5 overlap threshold.Unassigned priors receive no coordinate or class-prediction loss.Source: §4.5 · p. 13
5. Experiments Results
The experiments evaluate table detection across detector families and backbones using AP, AR, and F1 over IoU thresholds. Cascade Mask R-CNN with ResNeXt-101-64x4d achieves F1 0.844 over IoU 50%:95%.
5.1. Dataset and Metrics performance
- TNCR includes five table classes, with each class split into 70% training, 15% validation, and 15% testing data.The dataset is described as supporting detection and classification, including tables with no rows and columns.Source: §5.1 · pp. 13–14, Table 1 · p. 14
- Evaluation uses COCO-style AP, AR, and F1 across different IoU thresholds, with a true positive requiring the full table region to be captured cleanly.The authors use IoU overlap to determine whether a detected table region is correct.Source: §5.1 · pp. 14–15
5.2. Experiment Settings
- All models use images scaled to 1300×1500, batch size 16, SGD with momentum 0.9, weight decay 0.0001, learning rate 0.02, and an FPN neck.The experiments were implemented with MMDetection and run on Google Colaboratory using Tesla V100-SXM GPUs, with additional runs on a machine with Tesla K20X GPUs.Source: §5.2 · pp. 15–16
5.3. Results
- F1 0.844 over IoU 50%:95% is the highest reported Cascade Mask R-CNN result, achieved with the ResNeXt-101-64x4d backbone.At 50% IoU, this backbone also has the highest precision (0.891) and recall (0.975) among the compared backbones.Source: §5.3 · p. 16, Table 2 · p. 16
- Cascade-RCNN with ResNeXt-101-64x4d achieves F1 0.841 over IoU 50%:95%, while a separate comparison reports F1 0.931 versus 0.877 and 0.742 for ResNeXt101, ResNet101, and ResNet50.ResNet-50 with a 1× learning-rate schedule is reported as the lowest-performing Cascade-RCNN variant across various IoUs.Source: §5.3 · pp. 16–17, Table 3 · p. 17
- Fast R-CNN reaches F1 0.804 over IoU 50%:95%, compared with 0.609 for Cascade RPN; their reported average recalls over IoU 50% are 0.994 and 0.962, respectively.Average recall is defined here as the mean of recalls over IoU thresholds from 0.5 to 0.95 in steps of 0.05.Source: §5.3 · p. 17, Table 4 · p. 18
- Hybrid Task Cascade achieves its highest F1, 0.840 over IoU 50%:95%, with ResNet-50 and a 1× learning-rate schedule; the authors report overfitting for both ResNeXt-101 variants.ResNet-101 improves F1 by 2.8% over ResNet-50 with a 20e schedule.Source: §5.3 · p. 18, Table 5 · p. 18
- YOLO performs poorly overall for table detection, with DarkNet-53 at scale 320 achieving F1 0.492 and F1 0.042 at IoU 95%.The authors state that YOLO is not suitable for table detection in these experiments.Source: §5.3 · p. 18
6. Conclusion and future work
TNCR is presented as a real-image dataset for table analysis and evaluated with multiple detection models. The conclusion reports model F1 scores and identifies dataset and model extensions as future work.
6. Conclusion and future work
- TNCR is a real-image dataset intended to support research in table detection, structure recognition, and classification for document analysis.Source: §6 · p. 18
- Models were evaluated across IoU thresholds from 50% to 95%, and the authors report table detection as more difficult than cell structure detection.They describe image-based deep learning for table detection and recognition as a promising research direction.Source: §6 · p. 19
- F1 scores were 0.844 for Cascade Mask R-CNN, 0.841 for Cascade R-CNN, 0.804 for Cascade RPN, 0.840 for Hybrid Task Cascade (HTC), and 0.492 for YOLO.Source: §6 · p. 19
- Future work includes adding real labeled images, improving detection of tables near other image elements, and balancing classification classes.Source: §6 · p. 19