Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/KiLJ4EdeN/train-tesseract4-persian
Train Tesseract 4 for persian
https://github.com/KiLJ4EdeN/train-tesseract4-persian
Last synced: 3 months ago
JSON representation
Train Tesseract 4 for persian
- Host: GitHub
- URL: https://github.com/KiLJ4EdeN/train-tesseract4-persian
- Owner: KiLJ4EdeN
- Created: 2024-05-09T19:35:42.000Z (6 months ago)
- Default Branch: main
- Last Pushed: 2024-05-09T19:36:29.000Z (6 months ago)
- Last Synced: 2024-05-10T11:23:13.076Z (6 months ago)
- Language: Python
- Size: 1000 Bytes
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# tess-train
python version 3.9.19
# deps
```bash
sudo apt-get install -y wget unzip bc vim libleptonica-dev
sudo apt-get install -y --reinstall make # 4.2+
sudo apt-get install -y g++ autoconf automake libtool pkg-config libpng-dev libjpeg8-dev libtiff5-dev libicu-dev libpango1.0-dev autoconf-archive
sudo apt-get install libgirepository1.0-dev
sudo apt-get install libcairo2-dev libjpeg-dev libgif-dev
wget https://github.com/tesseract-ocr/tesseract/archive/4.1.0.zip
unzip 4.1.0.zip
rm -rf 4.1.0.zip
cd tesseract-4.1.0/
./autogen.sh
./configure
make -j4
sudo make install
sudo ldconfig
sudo make training
sudo make training-install
cd /usr/local/share/tessdata
sudo wget https://github.com/tesseract-ocr/tessdata_best/raw/main/fas.traineddata
export TESSDATA_PREFIX=/usr/local/share/tessdata
pip install -r requirements.txt
```
# data
note that these are assumed to be word images, if they are not just detect and crop them.
```bash
data/
ground-truth/
{idx}.png|.tiff # contains image
{idx}.gt.txt # contains ocr label text
validation/
{idx}.png|.tiff
{idx}.gt.txt
```
# train
```bash
cd tesseract-4.1.0
git clone https://github.com/tesseract-ocr/tesstrain
cd ..
mkdir -p tesseract-4.1.0/tesstrain/data/parsi-ground-truth
rsync -a ./data/ground-truth/ tesseract-4.1.0/tesstrain/data/parsi-ground-truth/
cd tesseract-4.1.0/tesstrain
```
## page segmentation mode
Set correct psm according to tesseract for training:
PSM Page segmentation mode. Default: 13
```bash
user@Workstation:~$ tesseract --help-psm
Page segmentation modes:
0 Orientation and script detection (OSD) only.
1 Automatic page segmentation with OSD.
2 Automatic page segmentation, but no OSD, or OCR. (not implemented)
3 Fully automatic page segmentation, but no OSD. (Default)
4 Assume a single column of text of variable sizes.
5 Assume a single uniform block of vertically aligned text.
6 Assume a single uniform block of text.
7 Treat the image as a single text line.
8 Treat the image as a single word.
9 Treat the image as a single word in a circle.
10 Treat the image as a single character.
11 Sparse text. Find as much text as possible in no particular order.
12 Sparse text with OSD.
13 Raw line. Treat the image as a single text line,
bypassing hacks that are Tesseract-specific.
```
Based on the info we select 8 for single words.
```bash
sudo make training MODEL_NAME=parsi START_MODEL=fas PSM=8 TESSDATA=/usr/local/share/tessdata MAX_ITERATIONS=100000 EPOCHS=100000 -j16
sudo cp data/parsi.traineddata /usr/local/share/tessdata
```
output example:
```bash
At iteration 7569/99900/100784, Mean rms=0.379%, delta=0.929%, char train=2.211%, word train=4.16%, skip ratio=0.5%, New worst char error = 2.211 wrote checkpoint.
At iteration 7572/100000/100884, Mean rms=0.347%, delta=0.71%, char train=1.755%, word train=3.86%, skip ratio=0.5%, New worst char error = 1.755 wrote checkpoint.
Finished! Error rate = 1.07
lstmtraining \
--stop_training \
--continue_from data/parsi/checkpoints/parsi_checkpoint \
--traineddata data/parsi/parsi.traineddata \
--model_output data/parsi.traineddata
Loaded file data/parsi/checkpoints/parsi_checkpoint, unpacking...
```
# test
```bash
python infer.py parsi
```