Handwriting Recognition1999en

IAM Handwriting Database

13,353 handwritten text lines from 657 writers. Standard handwriting benchmark.

Samples:13,353
Metrics:cer, wer
Paper / Website
Current State of the Art

DTrOCR 105M

Unknown

2.38

cer

cer Progress Over Time

Showing 5 breakthroughs from Sep 2018 to Aug 2023

2.03.24.45.66.8Sep 2018Nov 2019Feb 2021May 2022Aug 2023cerDate

Key Milestones

Sep 2018
Start, Follow, Read

From paper: Start, Follow, Read: End-to-End Full-Page Handwriting Recognition

6.4
May 2020
Transformer w/ CNN (+synth)

From paper: Pay Attention to What You Read: Non-recurrent Handwritten Text-Line Recognition

4.7
-27.0%
Dec 2020
VAN

From paper: End-to-end Handwritten Paragraph Text Recognition Using a Vertical Attention Network

4.3
-7.5%
Apr 2021
Self-Attention + CTC + language model

From paper: Rethinking Text Line Recognition Models

2.8
-36.3%
Aug 2023
DTrOCR 105MCurrent SOTA

From paper: DTrOCR: Decoder-only Transformer for Optical Character Recognition

2.4
-13.5%
Total Improvement
62.8%
Time Span
5y
Breakthroughs
5
Current SOTA
2.4

Top Models Performance Comparison

Top 10 models ranked by cer (lower is better)

cer1DTrOCR 105M2.4100.0%2 Self-Attention + CTC + l...2.886.5%3TrOCR-large 558M2.982.4%4Transformer + CNN3.080.4%5TrOCR-base 334M3.469.6%6TrOCR-small 62M4.256.4%7VAN4.355.1%8Transformer w/ CNN (+synth)4.751.0%9HTR-VT(line-level)4.750.6%10Easter2.06.238.3%0%25%50%75%100%% of best
Best Score
2.4
Top Model
DTrOCR 105M
Models Compared
10
Score Range
3.8

cerPrimary

#ModelScorePaper / CodeDate
1
DTrOCR 105M
2.38Aug 2023
2
Self-Attention + CTC + language model
2.75Apr 2021
3
TrOCR-large 558M
2.89Sep 2021
4
Transformer + CNN
2.96Apr 2021
5
TrOCR-base 334M
3.42Sep 2021
6
TrOCR-small 62M
4.22Sep 2021
7
VAN
4.32Dec 2020
8
Transformer w/ CNN (+synth)
4.67May 2020
9
HTR-VT(line-level)
4.7Sep 2024
10
Easter2.0
6.21May 2022
11
FPHR+Aug Paragraph Level (~145 dpi)
6.3Mar 2021
12
Start, Follow, Read
6.4
Start, Follow, Read: End-to-End Full-Page Handwriting RecognitionCode
Sep 2018
13
Decouple Attention Network
6.4Dec 2019
14
FPHR+Aug Line Level (~145 dpi)
6.5Mar 2021
15
Leaky LP Cell
6.6Feb 2019
16
FPHR Paragraph Level (~145 dpi)
6.7Mar 2021
17
Transformer w/ CNN
7.62May 2020

wer

#ModelScorePaper / CodeDate
1
HTR-VT(line-level)
14.9Sep 2024
2
Leaky LP Cell
15.9Feb 2019
3
VAN
16.24Dec 2020
4
Decouple Attention Network
19.6Dec 2019
5
Start, Follow, Read
23.2
Start, Follow, Read: End-to-End Full-Page Handwriting RecognitionCode
Sep 2018

Related Papers10

TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models
Sep 2021Models: TrOCR-small 62M, TrOCR-base 334M, TrOCR-large 558M
Rethinking Text Line Recognition Models
Apr 2021Models: Transformer + CNN, Self-Attention + CTC + language model
Full Page Handwriting Recognition via Image to Sequence Extraction
Mar 2021Models: FPHR Paragraph Level (~145 dpi), FPHR+Aug Line Level (~145 dpi), FPHR+Aug Paragraph Level (~145 dpi)
Pay Attention to What You Read: Non-recurrent Handwritten Text-Line Recognition
May 2020Models: Transformer w/ CNN, Transformer w/ CNN (+synth)
Decoupled Attention Network for Text Recognition
Dec 2019Models: Decouple Attention Network

Other Handwriting Recognition Datasets