conda create -n ahdoc python=3.10
conda activate ahdoc
conda install ipython pip
# python dependencies
pip install -r requirement.txt
# clone repo
git clone https://github.com/Hedrax/AHDoc.git
cd AHDoc/
- Datasets must follow the format of the custom Data format provided above
utils.pyhas all functions needed to migrate from JSON with Pascal and txt with Yolo formatutils.pyalso has all functions needed to list image names and get corresponding label files to list in .txt file
- Datasets must follow the format all images in png in the same directory with
labels.txtlabels must be in the formimageNameWithoutExtension_groundTruth utils.pyhas all functions needed for dataset preparation
- Follow instructions in
train.py - All training configurations are saved in
config.py
- put the weights files downloaded from the above reference to
./Text Detection/weights/and./OCR Engine/weights/ - Follow instructions in
inference.pyandevaluation.pyin terms of OCR Engine - All test configurations are saved in
config.py
We compare best-weights of universal model performance on our custom evaluation Arabic handwritten data
| Weights | Precision (%) | Recall (%) | F-measure (%) |
|---|---|---|---|
| Universal Model | 61.53 | 34.60 | 41.33 |
| Our-Model | 81.66 | 78.82 | 79.07 |
our results on the TEST set of 18-fonts
| # | Number of Words | Solid Accuracy% | Salted Accuracy% | Bolded Accuracy% | Notes | |||
|---|---|---|---|---|---|---|---|---|
| CRR | WRR | CRR | WRR | CRR | WRR | |||
| 1 | 1 | 94.28 | 70.05 | 91.85 | 57.08 | 77.25 | 17.81 | Tested on 7-Character Words |
| 2 | 1 | 94.24 | 54.06 | 91.42 | 50.94 | 90.19 | 46.04 | - |
| 3 | 2 | 89.81 | 39.38 | 87.11 | 34.84 | 86.75 | 33.85 | - |
| 4 | 3 | 89.64 | 35.63 | 88.23 | 37.39 | 87.79 | 35.59 | - |
| 5 | 4 | 82.23 | 28.59 | 80.41 | 24.61 | 80.25 | 24.61 | - |
| 6 | 5 | 73.17 | 20.25 | 71.88 | 17.225 | 70.25 | 16.62 | - |
| 7 | 6 | 66.01 | 18.78 | 64.95 | 13.48 | 63.50 | 14.39 | - |
