SLANet-1M: A Lightweight Model for Table Recognition

---

🧾 Overview

SLANet-1M is a lightweight convolutional model for table recognition designed to extract table structure and cell content from document images efficiently.
It is trained on over one million synthetic and real-world tables and provides competitive performance compared to transformer-based architectures—while being significantly smaller and faster.

This model was developed as part of a Master’s thesis at the University of Florence and the Swiss AI Center (iCoSys, Fribourg), and presented at SwissText 2025.

The paper is available here.

---

🚀 Key Features

• Lightweight architecture (≈9.2M parameters)
• Transformer-free design for CPU-friendly deployment
• Trained on large-scale datasets (PubTabNet + SynthTabNet)
• Compatible with deployment pipelines such as the Core Engine
• Outputs table structure in HTML format

---

📦 Model Details

Property	Description
Model Name	SLANet-1M
Architecture	CNN-based (SLANet variant with depthwise separable convolutions)
Parameters	~9.2 million
Input Size	480×480 (RGB)
Output Format	HTML table structure
Training Data	PubTabNet + SynthTabNet (all subsets)
Metrics	S-TEDS: 99.36 on SynthTabNet and 97.36 on PubTabNet

---

Please cite us:

@inproceedings{romaric-etal-2025-slanet,
    title = "{SLAN}et-1{M}: A Lightweight and Efficient Model for Table Recognition with Minimal Computational Cost",
    author = "Romaric, Nguinwa Mbakop Dimitri  and
      Petrucci, Andrea  and
      Marinai, Simone  and
      Hennebert, Jean",
    editor = {Gerber, Jonathan  and
      Cieliebak, Mark  and
      Tuggener, Don  and
      H{\"u}rlimann, Manuela},
    booktitle = "Proceedings of the 10th edition of the Swiss Text Analytics Conference",
    month = may,
    year = "2025",
    address = "Winterthur, Switzerland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.swisstext-1.9/",
    pages = "89--102"
}