Overview of Table Recognition Techniques and Practical Implementation

Tables are a common way to present structured data, and digitizing them from scanned documents or images often requires more than simple OCR because the table layout must be reconstructed. Two main OCR paradigms exist: a two‑stage pipeline (detect text lines then recognize them) and an end‑to‑end approach that predicts both location and content directly. While two‑stage OCR is more mature, table recognition also follows these two paradigms.

Four notable table‑recognition solutions are examined:

SPLERGE – a two‑stage method that first predicts horizontal and vertical lines to split the table into fine‑grained cells, then merges cells using a rule‑based or learned strategy. It handles both ruled and rule‑less tables but struggles with tilted tables and complex merges.

CascadeTabNet – an end‑to‑end model derived from CascadeRCNN, using HRNet as backbone to preserve details. It predicts whole‑table proposals, then refines cell locations via bounding‑box regression on text regions, achieving top scores on ICDAR‑2019 but requiring heavy GPU resources.

TableMASTER – adapts the MASTER OCR model to output HTML sequences representing table structure. It treats table reconstruction as a sequence‑to‑sequence task, predicting tags and bounding boxes for text cells, and performs well on the PubTabNet dataset.

UnetTable – combines a lightweight MobileNet backbone with a UNet head to detect visible table lines. It focuses on line detection and relies on traditional CV post‑processing to rebuild the table, offering a balance between simplicity and performance.

Based on these methods and specific business requirements, the article describes a practical deployment that adopts the UnetTable pipeline, adding pre‑table detection, handling of multi‑table images, and specialized post‑processing for both ruled and rule‑less tables. Enhancements such as a U2‑Net+CBAM backbone, removal of invisible‑line branches, and integration of YOLOX for rule‑less tables are detailed.

Evaluation metrics include Table Structure F1 (IoU ≥ 0.9 for cell alignment) and Table Structure + Text F1 (requiring both cell alignment and exact OCR text match). The presented solution surpasses major OCR vendors on a test set covering over 70 real‑world scenarios.

The article concludes that table recognition remains an open research problem, but the described approaches provide effective solutions for industrial applications.