Arabic-HLR-Model/README.md

metadata

language:
  - ar
license: cc-by-nc-4.0
tags:
  - handwritten-text-recognition
  - arabic
  - khatt
  - densenet
  - transformer
  - transfer-learning
  - pytorch
  - safetensors
datasets:
  - KHATT
  - DASTNUS
metrics:
  - cer
  - wer
pipeline_tag: image-to-text

Arabic Handwritten Text Recognition: DenseNet121-Transformer (Fine-tuned on KHATT)

Model Description

A lightweight DenseNet121-Transformer architecture for Arabic handwritten line recognition, pre-trained on the Kurdish DASTNUS dataset and fine-tuned on the KHATT Arabic handwritten dataset. Uses a triple unified vocabulary covering Kurdish, Arabic, and Urdu scripts (192 tokens). The KHATT dataset is publicly available at https://www.kaggle.com/datasets/iraqyomar/khatt-arabic-hand-written-lines/code (We only used Unique Handwritten Lines)

Architecture

• CNN Backbone: DenseNet-121 (pretrained on ImageNet)
• Encoder: 3 Transformer encoder layers
• Decoder: 3 Transformer decoder layers
• Attention Heads: 8
• Hidden Size: 256
• Parameters: ~12.8M
• Vocabulary: 192 tokens (Triple unified: Kurdish + Arabic + Urdu)

Transfer Learning Pipeline

1. Pre-trained on Kurdish DASTNUS dataset (with unified vocabulary)
2. Fine-tuned on KHATT Arabic handwritten line dataset

Performance on KHATT Test Set

Metric	Value
CER	0.1135
WER	0.4156
CRR	88.65%

Training Data

• Pre-training: DASTNUS Kurdish handwritten dataset
• Fine-tuning: KHATT Arabic handwritten dataset (5,166 training, 574 validation)

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", encoding="utf-8") as f:
    vocab = json.load(f)

# Load full unified vocabulary info
with open("unified_vocabulary.json", "r", encoding="utf-8") as f:
    unified_vocab = json.load(f)

Citation

[]

License

This model is released for non-commercial scientific research purposes only.