SimpleHTR — Word-Level Handwritten Text Recognition
A word-recognition model that reads individual handwritten words from cropped images. It takes a single word image and predicts the text. Part of the Xournal++ HTR project.
Model details
| Property | Value |
|---|---|
| Architecture | 5 CNN layers + 2 bidirectional LSTM layers + CTC output |
| Input | Grayscale image, resized to 128×32 (uniform scale, centered) |
| Output | CTC log-probabilities (seq_len × num_classes) |
| Decoding | Greedy best-path |
| Format | ONNX (opset 17) |
| Parameters | ~3.2M |
| Training data | IAM Handwriting Database |
| Best val CER | 0.056 (5.6%) |
| Best val word accuracy | 84.2% |
| License | MIT |
Usage
from xournalpp_htr.inference_models import SimpleHTRModel
model = SimpleHTRModel.from_pretrained()
text = model.recognize(word_image_grayscale) # str
Requires pip install xournalpp-htr (pulls onnxruntime and
huggingface-hub, no PyTorch needed).
How it works
The model processes a grayscale word image through:
- CNN encoder (5 layers): extracts visual features, reducing height to 1 and width to 32 timesteps via batch norm + ReLU + max pooling.
- Bidirectional LSTM (2 layers, hidden=256): models sequential dependencies across timesteps.
- CTC output layer: projects to character probabilities at each timestep.
- Greedy decoding: collapses repeated characters and removes blanks.
Input images are resized with uniform scaling (preserving aspect ratio) and centered on a white 128×32 canvas, then normalised to [-0.5, 0.5].
Training
Trained on the IAM Handwriting Database (word-level, 95/5 random split). Best configuration found via grid search over 3 experiments:
| Hyperparameter | Value |
|---|---|
| Optimizer | Adam |
| Learning rate | 0.0005 |
| Batch size | 64 |
| Dropout | 0.5 |
| Data augmentation | Enabled (blur, geometric, morphological, contrast) |
| Max epochs | 200 |
| Early stopping patience | 25 |
Full training instructions and experiment logs: docs/models/simple_htr.md.
Intended use
This model is the recognition stage in a handwriting recognition pipeline, following a word detector that crops individual words. It is designed to run on personal devices (laptops, edge) via ONNX Runtime — no GPU required for inference.
Limitations
- Single-word recognition only — expects a pre-cropped word image.
- Greedy CTC decoding (no language model); beam search could improve results.
- Grayscale input required.
- Trained on IAM handwriting; performance on other styles may vary.
- 78-character charset (English letters, digits, punctuation) — does not support non-Latin scripts.
Citation
The architecture is based on SimpleHTR by Harald Scheidl.
- Downloads last month
- 7