Handwritten text recognition on IAM (test)
3.2CERBluche et al.
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Bluche et al.System=RNN+CTC, Lexicon size (k)=502020.05 | 3.2 | 10.5 | — | |
| TrOCRextra training data=true, Parameters=385.0M2024.09 | 3.4 | — | — | |
| HTR-ConvTextParam.=65.9M2025.12 | 4 | 12.9 | — | |
| [7]2021.02 | 4.32 | 16.24 | — | |
| PuigcerverSystem=RNN+CTC, Lexicon size (k)=502020.05 | 4.4 | 12.2 | — | |
| Transformer-based HTRSystem=Transf., Lexicon size (k)=none2020.05 | 4.67 | 15.45 | — | |
| Doetsch et al.System=HMM/ANN, Lexicon size (k)=502020.05 | 4.7 | 12.2 | — | |
| GTR-12 OrigamiNetProcessing Level=Multi-line, Input Scale=750 x 7502020.06 | 4.7 | — | — | |
| TransformersParameters=54.7M2024.09 | 4.7 | 15.5 | — | |
| HTR-VTParameters=53.5M2024.09 | 4.7 | 14.9 | — | |
| TransformerParam.=54.7M, Extra training data=true2025.12 | 4.7 | 15.5 | — | |
| HTR-VTParam.=53.5M2025.12 | 4.7 | 14.9 | — | |
| [22]2021.02 | 4.7 | — | — | |
| OrigamiNet-18Parameters=77.1M2024.09 | 4.8 | — | — | |
| OrigamiNet-24Parameters=115.3M2024.09 | 4.8 | — | — | |
| OrigamiNet-18Param.=77.1M2025.12 | 4.8 | — | — | |
| OrigamiNet-24Param.=115.3M2025.12 | 4.8 | — | — | |
| [35]Processing Level=Single-line, Input Scale=32 x W, Architecture=CNN+CTC2020.06 | 4.9 | — | — | |
| VANParameters=2.7M2024.09 | 5 | 16.3 | — | |
| VANParam.=2.7M2025.12 | 5 | 16.3 | — | |
| Kozielski et al.System=HMM/ANN, Lexicon size (k)=502020.05 | 5.1 | 13.3 | — | |
| Pham et al.System=RNN+CTC, Lexicon size (k)=502020.05 | 5.1 | 13.6 | — | |
| [18]Processing Level=Single-line, Input Scale=64 x W, Architecture=Seq2Seq (CNN+BLSTM encoder)2020.06 | 5.24 | — | — | |
| OrigamiNet-12Parameters=39.0M2024.09 | 5.3 | — | — | |
| OrigamiNet-12Param.=39.0M2025.12 | 5.3 | — | — | |
| SPAN2021.02 | 5.45 | 19.83 | — | |
| BlucheSystem=Seq2Seq, Lexicon size (k)=502020.05 | 5.5 | 16.4 | — | |
| GTR-8 OrigamiNetProcessing Level=Multi-line, Input Scale=750 x 7502020.06 | 5.5 | — | — | |
| GTR-8 OrigamiNetProcessing Level=Multi-line, Input Scale=500 x 5002020.06 | 5.6 | — | — | |
| Dutta et al.System=RNN+CTC, Lexicon size (k)=none2020.05 | 5.7 | 17.82 | — | |
| PuigcerverSystem=RNN+CTC, Lexicon size (k)=none2020.05 | 5.8 | 18.4 | — | |
| [22]Processing Level=Single-line, Input Scale=128 x W, Architecture=CNN+BLSTM+CTC2020.06 | 5.8 | — | — | |
| OrigamiNet-12re-implementation=by LAM [34], Parameters=39.0M2024.09 | 6 | 22.3 | — | |
| OrigamiNet-12Param.=39.0M, Implementation source=re-implemented by LAM2025.12 | 6 | 22.3 | — | |
| ResNet-74 OrigamiNetProcessing Level=Multi-line, Input Scale=500 x 5002020.06 | 6.1 | — | — | |
| OrigamiNetYear=2020, Backbone=ResNet2021.06 | 6.1 | — | — | |
| ExCTCYear=20202021.06 | 6.16 | — | — | |
| DANRecognition type=Line-level, Input=Text lines2019.12 | 6.4 | 19.6 | — | |
| Wigington et al.System=RNN+CTC, Lexicon size (k)=none2020.05 | 6.4 | 23.2 | — | |
| [33]Processing Level=Multi-line, Remarks=Requires full line-break annotation and partial visual localization2020.06 | 6.4 | — | — | |
| SFRYear=20182021.06 | 6.4 | — | — | |
| [21]2021.02 | 6.4 | 23.2 | — | |
| OrigamiNet-24re-implementation=by LAM [34], Parameters=115.3M2024.09 | 6.5 | 23.9 | — | |
| OrigamiNet-24Param.=115.3M, Implementation source=re-implemented by LAM2025.12 | 6.5 | 23.9 | — | |
| OrigamiNet-18re-implementation=by LAM [34], Parameters=77.1M2024.09 | 6.6 | 24.2 | — | |
| OrigamiNet-18Param.=77.1M, Implementation source=re-implemented by LAM2025.12 | 6.6 | 24.2 | — | |
| España et al.System=HMM/ANN, Lexicon size (k)=52020.05 | 6.9 | 15.5 | — | |
| TrOCRre-implementation=by LAM [34], Parameters=385.0M2024.09 | 7.3 | 37.5 | — | |
| TrOCRParam.=385.0M, Implementation source=re-implemented by LAM2025.12 | 7.3 | 37.5 | — | |
| Pastor et al.System=HMM/ANN, Lexicon size (k)=1032020.05 | 7.5 | 19 | — | |
| Zamora et al.System=HMM/ANN, Lexicon size (k)=1032020.05 | 7.6 | 16.1 | — | |
| Transformerre-implementation=by [46], Parameters=54.7M2024.09 | 7.6 | 24.5 | — | |
| TransformerParam.=54.7M, Implementation source=re-implemented by LAM2025.12 | 7.6 | 24.5 | — | |
| CNN + BLSTMre-implementation=by LAM [34], Parameters=9.3M2024.09 | 7.7 | 26.3 | — | |
| CNN + BLSTMParam.=9.3M, Implementation source=re-implemented by LAM2025.12 | 7.7 | 26.3 | — | |
| CNN + 1D-LSTMArchitecture=CNN + 1D-LSTM [6]2020.12 | 7.73 | 25.22 | — | |
| CRNNre-implementation=by LAM [34], Parameters=18.2M2024.09 | 7.8 | 27.8 | — | |
| CRNNParam.=18.2M, Implementation source=re-implemented by LAM2025.12 | 7.8 | 27.8 | — | |
| Bluche (2016)2019.12 | 7.9 | 24.6 | — | |
| BlucheSystem=Seq2Seq, Lexicon size (k)=none2020.05 | 7.9 | 24.6 | — | |
| [3]Processing Level=Multi-line, Input Scale=300 dpi2020.06 | 7.9 | — | — | |
| BlucheYear=20162021.06 | 7.9 | — | — | |
| [1]2021.02 | 7.9 | 24.6 | — | |
| GFCNArchitecture=Ours2020.12 | 7.99 | 28.61 | — | |
| GFCNParameters=1.4M2024.09 | 8 | 28.6 | — | |
| GFCNre-implementation=by LAM [34], Parameters=1.4M2024.09 | 8 | 28.6 | — | |
| GFCNParam.=1.4M2025.12 | 8 | 28.6 | — | |
| GFCNParam.=1.4M, Implementation source=re-implemented by LAM2025.12 | 8 | 28.6 | — | |
| ChowdhurySystem=Seq2Seq, Lexicon size (k)=none2020.05 | 8.1 | 16.7 | — | |
| CNN + 1D-LSTMArchitecture=CNN + 1D-LSTM [5]2020.12 | 8.2 | — | — | |
| CNN + BLSTMParameters=9.3M2024.09 | 8.3 | 24.9 | — | |
| CNN + BLSTMParam.=9.3M2025.12 | 8.3 | 24.9 | — | |
| Bhunia et al. (2019)Recognition type=Word-level, Input=Cropped word images2019.12 | 8.4 | 17.2 | — | |
| Zhang et al. (2019)2019.12 | 8.5 | 22.2 | — | |
| [7]Processing Level=Multi-line, Remarks=Requires fully segmented training data2020.06 | 8.5 | — | — | |
| [6]2021.02 | 8.5 | — | — | |
| Sueiras et al. (2018)2019.12 | 8.8 | 23.8 | — | |
| 2D-LSTM-X2Architecture=2D-LSTM-X2 [6]2020.12 | 8.86 | 29.31 | — | |
| 2D-LSTMArchitecture=2D-LSTM [6]2020.12 | 8.88 | 29.15 | — | |
| SpiS-GANInput height=64-pixel, variant=Ours-642026.07 | 9.01 | 26.02 | 8.83 | |
| HiGAN+Input height=64-pixel2026.07 | 9.45 | 27.33 | 9.58 | |
| Krishnan et al.System=RNN+CTC, Lexicon size (k)=none2020.05 | 9.78 | 32.89 | — | |
| Salvador et al. (2011)2019.12 | 9.8 | 22.4 | — | |
| España et al.System=HMM/ANN, Lexicon size (k)=none2020.05 | 9.8 | 22.4 | — | |
| ADPYear=20202021.06 | 9.97 | — | — | |
| SpiS-GANInput height=32-pixel, variant=Ours-322026.07 | 10.03 | 27.71 | 9.99 | |
| [3]Processing Level=Multi-line, Input Scale=150 dpi2020.06 | 10.1 | — | — | |
| FWGANInput height=32-pixel2026.07 | 10.18 | 28.38 | 10.19 | |
| Dreuw et al.System=HMM/ANN, Lexicon size (k)=502020.05 | 10.3 | 29.2 | — | |
| VATrInput height=32-pixel2026.07 | 10.52 | 29.15 | 10.32 | |
| HWTInput height=32-pixel2026.07 | 10.75 | 29.31 | 10.7 | |
| One-DMInput height=64-pixel2026.07 | 10.76 | 30.26 | 10.65 | |
| Pham et al. (2014)2019.12 | 10.8 | 35.1 | — | |
| Pham et al.System=RNN+CTC, Lexicon size (k)=none2020.05 | 10.8 | 35.1 | — | |
| 5000 real imagesInput height=64-pixel2026.07 | 10.81 | 30.31 | 10.66 | |
| HiGANInput height=32-pixel2026.07 | 10.98 | 30.02 | 10.51 | |
| Chen et al.System=RNN+CTC, Lexicon size (k)=none2020.05 | 11.15 | 34.55 | — | |
| Almazán et al.System=HMM/ANN, Lexicon size (k)=none2020.05 | 11.27 | 20.01 | — | |
| DiffusionPenInput height=64-pixel2026.07 | 11.86 | 31.81 | 11.52 | |
| 5000 real imagesInput height=32-pixel2026.07 | 12.25 | 32.81 | 11.95 |