tesseract-languages

ESB.AI Lab repo of released trained tesseract-ocr languages for language preservation

Welcome to our new Hawaiian Model! We trained this using tesseract-ocr on 2 books with 234 pages total - O Kamehameha (136 pages) and O Lunalilo (98 pages). This resulted in a BCER of 1.275% after 20K Iterations.

See below for how to replicate training our Hawaiian model with our labeled data, as well as the results below!

Hawaiian Language Notes

Hawaiian Alphabet

The Hawaiian alphabet consists of 13 letters: H, K, L, M, N, P, W, A, E, I, O, U, ʻ (ʻokina) and with vowels that can contain kahako's (ā, ē, ī, ō, ū).

Data scraping, data labeling and model training done by:

• Edwin Solares, Ph.D. Lead PI
• Anthony Tong
• Welo Gosline-Niheu
• Andrew Pan
• Sadrac Santa-Cruz
• Doanh Nguyen
• Yashil Vora
• Nikhil Rao, M.S.

We used LabelStudio community edition and Docker for assisting in data labeling and Singularity CE and the San Diego Super Computing Cluster for training.

---

📌 How to Train Tesseract OCR for Hawaiian

🔹 1. Install Tesseract & Tesstrain

First, install you should install Tesseract on your machine, using the installation guide here.

Verify the installation:

$ tesseract --version

Then, you should clone Tesstrain:

$ git clone https://github.com/tesseract-ocr/tesstrain.git

Then, navigate to tesstrain and make a data directory in it like tesstrain/data, i.e.

$ cd tesstrain
$ mkdir -p data

We also need to download the existing trained languages:

$ git clone https://github.com/tesseract-ocr/langdata_lstm.git data/langdata

---

🔹 2. Prepare Training Files & Ground Truth Data

Our training data located in the repo is hawaiian-ground-truth, which includes .png files and .gt.txt each having a line of Hawaiian text and its corresponding ground truth. To train, we need to move the ground truth data into the tesstrain directory, specifically in tesstrain/data/hawaiian-ground-truth/.

If your current working directory is our repo you've cloned, you would move the directory hawaiian-ground-truth to tesstrain/data/hawaiian-ground-truth/ making sure to modify the tesstrain path to wherever it is on your computer. For example:

$ cp -r hawaiian-ground-truth ./tesstrain/data/hawaiian-ground-truth/

Now we have our Hawaiian training data ready!

---

🔹 3. Create Required Hawaiian Files

Tesseract needs additional files in the tesstrain directory:

$ mkdir -p tesstrain/data/langdata/hawaiian 
$ mkdir -p tesstrain/data/mri
$ wget -P tesstrain/data/https://github.com/tesseract-ocr/tessdata_best/raw/main/mri.traineddata

Also, there is a hawaiian folder in this repo that contains the language details of Hawaiian that Tesseract needs to train, including the Hawaiian vocabulary, punctuation, and numbers. You should copy over the hawaiian directory from this repo to your tesstrain/data/hawaiian directory, i.e.

$ cp -r hawaiian tesstrain/data/hawaiian

🔹 4. Train Tesseract on Hawaiian

Now we can start training the Hawaiian language for Tesseract!

From tesstrain, you can start running the model by:

$ make training MODEL_NAME=hawaiian START_MODEL=mri PSM=6 TESSDATA=data MAX_ITERATIONS=20000

This might take around 20-30 minutes, depending on your computer. It automatically saves checkpoints.

If you need to stop:

$ CTRL + C

Resume later with:

$ lstmtraining \
  --continue_from data/hawaiian/checkpoints/hawaiian_checkpoint \
  --traineddata data/hawaiian/hawaiian.traineddata \
  --model_output data/hawaiian/checkpoints/hawaiian \
  --train_listfile data/hawaiian/list.train \
  --eval_listfile data/hawaiian/list.eval \
  --max_iterations 20000 \
  --target_error_rate 0.01

---

🔹 5. Finalize the Model

Once training finishes:

$ lstmtraining \
  --stop_training \
  --continue_from data/hawaiian/checkpoints/hawaiian_best_checkpoint \
  --traineddata data/hawaiian/hawaiian.traineddata \
  --model_output data/hawaiian.traineddata

You can move the trained model to Tesseract, wherever your tessdata of all the languages are stored in tesseract:

$ sudo cp tesstrain/data/hawaiian.traineddata tessdata

Verify the model:

$ tesseract --list-langs

You should see hawaiian listed.

---

Results & Graphs

We evaluated our Hawaiian Tesseract model performance based on the Between Character Error Rate (BCER) over the 20K training iterations ran, following the instructions here. We end up achieving a training BCER of 1.275%, and validation BCER of 1.810%! See the results below:

🔹 Test Your Own Hawaiian Images

Now, you can run Tesseract OCR with Hawaiian on a test image:

$ tesseract test_image.png output -l hawaiian
$ cat output.txt

For example, see the following images from hawaiian-test-images and their Tesseract predictions in hawaiian-images-results as shown below!

test1

test2

test3

test4