metadata
language:
- ar
license: cc-by-nc-4.0
tags:
- handwritten-text-recognition
- paragraph-recognition
- ar
- densenet
- transformer
- pytorch
- safetensors
datasets:
- KHATT
metrics:
- cer
- wer
pipeline_tag: image-to-text
KHATT-Arabic: DenseNet121-Transformer Paragraph HTR
Model Description
Arabic handwritten paragraph recognition model evaluated on the KHATT dataset for cross-script generalisation accessed through: https://www.kaggle.com/datasets/iraqyomar/khatt-arabic-hand-written-lines/code. Pre-trained on 12,000 synthetic paragraphs combining KHATT Arabic lines with Kurdish lines from DASTNUS, then fine-tuned on 1,193 reconstructed KHATT paragraphs. Achieves CER of 0.1394, surpassing a reimplemented state-of-the-art baseline under identical conditions.
Architecture
- CNN Backbone: DenseNet-121 (pretrained on ImageNet)
- Horizontal Upsample: Yes
- Encoder: 3 Transformer encoder layers
- Decoder: 6 Transformer decoder layers
- Attention Heads: 8
- Hidden Size: 256
- Feed-Forward Dim: 2048
- Vocabulary Size: 143
- Parameters: 22,760,778
Performance on KHATT
| Metric | Value |
|---|---|
| CER (greedy) | 0.1394 |
| WER (greedy) | 0.5075 |
Input Format
- Image size: 600 x 1235 pixels
- Preprocessing: Aspect-ratio-preserving resize, right-aligned on white canvas (RTL)
- Normalization: ImageNet mean/std
Training
- Pre-training: 12,000 synthetic paragraph images with curriculum learning
- Fine-tuning: Real handwritten paragraphs from KHATT
- Two-stage strategy: Encoder frozen for first 10 epochs during fine-tuning
Usage
from safetensors.torch import load_file
import json
# Load model weights
state_dict = load_file("model.safetensors")
# Load config
with open("config.json", "r") as f:
config = json.load(f)
# Load vocabulary
with open("vocab.json", "r") as f:
vocab = json.load(f)
# Load reverse mapping
with open("idx_to_char.json", "r") as f:
idx_to_char = json.load(f)
Citation
[Citation to be added upon publication]
License
This model is released under CC-BY-NC-4.0 for non-commercial research purposes only.