Models

Scene text detection models, compared.

Newest first. One table, one vocabulary: each row links to a page with the paper, the idea in plain words, results with sources, and where the code lives.

ModelYearFamilyOutputReported resultsSpeed (paper)Code
PP-OCR det
Runs on this site
2026Segmentation (DB-family, mobile backbones)Quadrilaterals per text linenone stated; see pageTens of milliseconds per image on WebGPU (measured live in the detector's status line)PaddlePaddle/PaddleOCR
FAST
2021Segmentation (single-channel kernel + GPU dilation)Polygons of arbitrary shape
Total-Text F-measure 0.816 · FAST-T at 152 FPS (abstract)
152 FPS for FAST-T on Total-Text; over 600 FPS with TensorRT (abstract)czczup/FAST
DBNet
Runs on this site
2020Segmentation (differentiable binarization)Polygons of arbitrary shape
MSRA-TD500 F-measure 0.828 · ResNet-18 backbone at 62 FPS (abstract)
ICDAR 2015 F-measure 0.873 · ResNet-50 backbone; table value, not yet checked against the PDFunverified
62 FPS with ResNet-18 on MSRA-TD500 (abstract)MhLiao/DB
CRAFT
2019Segmentation (character regions + affinity)Polygons around words, built from character-level maps
ICDAR 2015 F-measure 0.869 · table value; the abstract states no numbersunverified
clovaai/CRAFT-pytorch
EAST
2017Direct regression (fully convolutional, one stage)Rotated rectangles or quadrilaterals per word
ICDAR 2015 F-score 0.782 · at 13.2 FPS, 720p input (abstract)
ICDAR 2015 F-score (best configuration) 0.807 · table value, not yet checked against the PDFunverified
13.2 FPS at 720p resolution (abstract)argman/EAST
CTPN
2016Proposal-based (vertical anchors + recurrent connection)Horizontal text lines
ICDAR 2013 F-measure 0.88 · abstract
ICDAR 2015 F-measure 0.61 · abstract
0.14 s per image with a VGG-16 backbone on a 2016 GPU (abstract)Original Caffe implementation

How to read this table

Results are the papers' own claims. A value marked unverified was quoted from a results table and has not yet been checked against the PDF; values without the mark come from the abstract. Blank beats wrong, so some rows are sparse.

Datasets differ. ICDAR 2013 is easy, ICDAR 2015 is hard, Total-Text and CTW1500 are curved, MSRA-TD500 is line-level. A 0.83 on one and a 0.61 on another can be the same quality of detector. The datasets pages explain what each one tests.

Speed is not comparable. GPUs, input sizes and what counts as "inference" (with or without post-processing) changed between 2016 and 2021. Treat the speed column as each paper's headline, not a benchmark.

Families in one line each

Proposal-based: anchors along a line, linked into text (CTPN).

Direct regression: geometry predicted per pixel, merged by NMS (EAST).

Segmentation: a per-pixel text map turned into shapes (CRAFT, DBNet, FAST, PP-OCR).

The full timeline →