Text detection#

Text detection models localize the words of a page. In docTR, a word is an uninterrupted sequence of characters. All models predict a segmentation map that is turned into boxes, and support both straight and rotated boxes (assume_straight_pages=False).

Architectures#

Family

Paper

Variants

DBNet

Real-time Scene Text Detection with Differentiable Binarization

db_resnet34(), db_resnet50(), db_mobilenet_v3_large()

LinkNet

LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation

linknet_resnet18(), linknet_resnet34(), linknet_resnet50()

FAST

FAST: Faster Arbitrarily-Shaped Text Detector with Minimalist Kernel Representation

fast_tiny(), fast_small(), fast_base() default

Benchmark#

Recall and precision on FUNSD and CORD, and latency with a batch size of 1:

Architecture

Input shape

# params

FUNSD recall

FUNSD precision

CORD recall

CORD precision

sec/it

db_resnet34

(1024, 1024, 3)

22.4 M

82.76

76.75

89.20

71.74

0.8

db_resnet50

(1024, 1024, 3)

25.4 M

83.56

86.68

92.61

86.39

1.1

db_mobilenet_v3_large

(1024, 1024, 3)

4.2 M

82.69

84.63

94.51

70.28

0.5

linknet_resnet18

(1024, 1024, 3)

11.5 M

81.64

85.52

88.92

82.74

0.6

linknet_resnet34

(1024, 1024, 3)

21.6 M

81.62

82.95

86.26

81.06

0.7

linknet_resnet50

(1024, 1024, 3)

28.8 M

81.78

82.47

87.29

85.54

1.0

fast_tiny

(1024, 1024, 3)

13.5 M (8.5 M)

84.90

85.04

93.73

76.26

0.7 (0.4)

fast_small

(1024, 1024, 3)

14.7 M (9.7 M)

85.36

86.68

94.09

78.53

0.7 (0.5)

fast_base

(1024, 1024, 3)

16.3 M (10.6 M)

84.95

86.73

94.39

85.36

0.8 (0.5)

Note

FAST models are reparameterized by detection_predictor and ocr_predictor: their branches are fused into single convolutions for inference. Values in parentheses are measured after reparameterization.

Usage#

from doctr.models import ocr_predictor, detection_predictor

# In the OCR pipeline
predictor = ocr_predictor(det_arch="db_resnet50", pretrained=True)

# On its own
det_predictor = detection_predictor("db_resnet50", pretrained=True)

Tune the binarization and box thresholds of the post-processor to detect more (or fewer) words, see Tune the text detection.