Table structure recognition

Table structure models parse the image of a table into a machine-understandable grid: they localize every cell (spatial structure) and recover the rows and columns it spans (logical structure).

Architectures

Architecture	Paper
tablecenternet() default	TableCenterNet: A one-stage network for table structure recognition

Benchmark

Architecture	Input shape	# params	Recall	Precision	F1	Structure acc.	sec/it (batch 1)
tablecenternet	(1024, 1024, 3)	7.1 M	82.31	96.01	88.64	77.53	0.5

The metrics are produced by references/table/evaluate.py with TableCellMetric: cell detection recall, precision and F1 (cells matched above an IoU of 0.5), and structure accuracy, the share of matched cells whose row and column spans are correct.

Usage

from doctr.models import ocr_predictor, table_predictor

# In the OCR pipeline: tables are located by the layout model, then parsed
predictor = ocr_predictor(pretrained=True, detect_tables=True)

# On its own, on images of tables
tables = table_predictor("tablecenternet", pretrained=True)

See Table structure recognition to work with the recognized tables.