Handwritten Text Recognition on IAM Aachen (test)
2.26CERDRetHTRBASE
Evaluation Results
| Method | Links | |
|---|---|---|
| DRetHTRBASEParams=107M, Encoder=–, Decoder=RetNet w/EN-V2, L12/h12/d768, Train=Syn+IAM, LM=No2026.02 | 2.26 | |
| DRetHTR_BASEParams=107M, Tokenization=Character Level, Total Tokens=442832, Inference Time=123s, Memory=22.14GB2026.02 | 2.26 | |
| DTrHTR_BASEParams=107M, Tokenization=Character Level, Total Tokens=442832, Inference Time=233s, Memory=36.25GB2026.02 | 2.35 | |
| DTrOCRParams=105M, Encoder=–, Decoder=GPT-2, L12/h12/d768, Train=Syn+IAM, LM=No2026.02 | 2.38 | |
| Diaz et al.Params=1.5M, Encoder=SelfAttn, L16/h4/d256, Decoder=CTC, Train=Int.+IAM, LM=Yes2026.02 | 2.75 | |
| TrOCRLARGEParams=558M, Encoder=BEiTLARGE, L24/h16/d1024, Decoder=RoBERTaLARGE, L12/h16/d1024, Train=Syn+IAM, LM=No2026.02 | 2.89 | |
| TrOCR_LARGEParams=558M, Tokenization=BPE(GPT-2) |V| = 50,257, Total Tokens=31959, Inference Time=2725s, Memory=13.5GB2026.02 | 2.89 | |
| DRetHTRSMALLParams=73M, Encoder=–, Decoder=RetNet w/EN-V2, L4/h8/d1024, Train=Syn+IAM, LM=No2026.02 | 2.97 | |
| DRetHTR_SMALLParams=73M, Tokenization=Character Level, Total Tokens=442832, Inference Time=67s, Memory=14.5GB2026.02 | 2.97 | |
| DTrHTR_SMALLParams=73M, Tokenization=Character Level, Total Tokens=442832, Inference Time=110s, Memory=25.1GB2026.02 | 2.98 | |
| DRetHTR_SMALLParams=74M, Tokenization=BPE |V| = 500, Total Tokens=49574, Inference Time=44s, Memory=14.9GB2026.02 | 3.01 | |
| DTrHTR_SMALLParams=74M, Tokenization=BPE |V| = 500, Total Tokens=49574, Inference Time=69s, Memory=23.9GB2026.02 | 3.01 | |
| Bluche and MessinaParams=0.75M, Encoder=GCRNN, 6Conv+2BiLSTM, d128, Decoder=CTC, Train=Syn+IAM, LM=No2026.02 | 3.2 | |
| TrOCR_BASEParams=334M, Tokenization=BPE(GPT-2) |V| = 50,257, Total Tokens=31959, Inference Time=1717s, Memory=12.3GB2026.02 | 3.42 | |
| TrOCR_SMALLParams=62M, Tokenization=BPE(GPT-2) |V| = 50,257, Total Tokens=31959, Inference Time=310s, Memory=9.1GB2026.02 | 4.22 | |
| DRetHTRSMALLParams=73M, Encoder=–, Decoder=RetNet w/EN-V2, L4/h8/d1024, Train=IAM, LM=No2026.02 | 4.35 | |
| PuigcerverParams=≈8.3M, Encoder=5Conv, Decoder=5×BiLSTM d256 + CTC, Train=IAM, LM=Yes2026.02 | 4.4 | |
| VANParams=2.7M, Encoder=18 VGG-style Conv, Decoder=1×LSTM d256 + CTC, Train=IAM, LM=No2026.02 | 4.45 | |
| Kang et al.Params=100M, Encoder=Tx w/ResNet50, L4/h8/d1024, Decoder=Tx, L4/h8/d1024, Train=Syn+IAM, LM=No2026.02 | 4.67 | |
| HTR-VTParams=53.5M, Encoder=ResNet18 + Tx-Enc, L4/h6/d768, Decoder=CTC, Train=IAM, LM=No2026.02 | 4.7 | |
| Michael et al.Params=≈1.8M, Encoder=3Conv+3BLSTM, d256, Decoder=Attn-LSTM, L1/d256, Train=IAM, LM=No2026.02 | 4.87 | |
| Wick et al.Params=≈22M, Encoder=3Conv + Tx, L3/h8/d512, Decoder=Bidi Tx, L3/h8/d512, Train=IAM, LM=No2026.02 | 5.67 | |
| DANParams=≈2.18M, Encoder=23Conv+GRU, Decoder=GRU, L1/d256, Train=IAM, LM=No2026.02 | 6.4 |