EasyOCR on GX10: Simple Cyrillic OCR

01What you will learn

• How EasyOCR works: one model finds where the text is, another reads it.
• How to put PyTorch for ARM64 with CUDA into a virtual environment and install EasyOCR.
• How to create a Reader for Bulgarian and English and reuse it.
• How to read the result: a box, the text and a confidence score.
• How to switch to the CPU or work without internet.
• When EasyOCR is not the right choice.

02Before you start

• A machine of the NVIDIA GB10 class (for example ASUS Ascent GX10 or DGX Spark) with DGX OS (Ubuntu 24.04 for Arm64).
• Access to the machine's terminal — directly or over SSH.
• Python 3 with the venv module. Ubuntu 24.04 ships Python 3.12 by default — check with python3 --version, because the PyTorch wheels below are for 3.12.
• Internet on the first run: PyTorch (more than a gigabyte), the libraries and the model weights are downloaded.
• Free disk space — check with df -h.

03Steps

1. What EasyOCR does

   According to the project, recognition runs in two stages. First the CRAFT model (Character Region Awareness For Text) finds the regions that contain text — according to its paper it is designed to detect rotated and curved text as well. Then the recognition model (CRNN) reads each region character by character. Everything runs on PyTorch. Over 80 languages are supported; Bulgarian has the code bg, English en. The library's licence is Apache-2.0.

   💡

   Which languages work together

   bg is in the project's Cyrillic group together with Russian (ru), Ukrainian (uk), Belarusian (be), Serbian in Cyrillic (rs_cyrillic) and others. According to the project's documentation, English is compatible with every language and languages with similar characters usually work together — but not every pair is allowed. Start with ['bg','en'].

2. Check the machine

   bash · on the machinecopy

   uname -m
   python3 --version
   nvidia-smi

   You expect aarch64, Python 3.12 and a table with the GPU. The memory row may say "Memory-Usage: Not Supported" — on GB10 that is normal: the GPU uses the machine's shared memory (confirmed by NVIDIA).

3. A virtual environment and PyTorch for aarch64

   Why a virtual environment? PyTorch and the EasyOCR dependencies are heavy, and you should not mix their versions with system packages. The environment is deleted with one command without touching the system.

   Why an explicit index? This way you choose the CUDA build for ARM64 yourself instead of hoping pip picks a suitable one. If python3 -m venv complains that the module is missing, install python3-venv with sudo apt install python3-venv.

   bash · on the machinecopy

   python3 -m venv ~/ocr-env
   source ~/ocr-env/bin/activate
   pip install --upgrade pip
   pip install torch torchvision --index-url https://download.pytorch.org/whl/cu130
   python -c "import torch; print(torch.__version__, torch.cuda.is_available())"

   The last line should print the version and True. If it prints False or a warning about the GPU's compute capability, stop: PyTorch will not use the GPU. Then either look for a newer build in the index or carry on with the CPU (see step 7).

   ⚠️

   We have not run this on GB10

   The existence of an aarch64 wheel with CUDA 13.0 is checked in the index, but whether it supports the GB10 GPU without a warning we have not tested. That is why the check above is mandatory.

4. Install EasyOCR

   PyTorch is already in place, so pip will not replace it. It brings the rest: OpenCV, SciPy, NumPy, Pillow, scikit-image and others.

   bash · inside the virtual environmentcopy

   pip install easyocr
   python -c "import easyocr; print('easyocr: OK')"

   You will get version 1.7.2 — the latest to date.

5. A first read of a synthetic image

   So that you need no one else's files, we create an image with Bulgarian text ourselves. We use the DejaVu Sans font, which has Cyrillic glyphs; check that it exists with fc-list | grep -i dejavu and fix the path on the first line if it differs.

   python · ocr_test.pycopy

   from PIL import Image, ImageDraw, ImageFont
   import easyocr
   
   FONT = "/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf"
   
   # 1) a synthetic image with Cyrillic text
   img = Image.new("RGB", (900, 220), "white")
   draw = ImageDraw.Draw(img)
   font = ImageFont.truetype(FONT, 56)
   draw.text((30, 20), "Добър ден, Варна", fill="black", font=font)
   draw.text((30, 110), "Фактура № 2026-0417", fill="black", font=font)
   img.save("test.png")
   
   # 2) a Reader for Bulgarian and English — created ONCE
   reader = easyocr.Reader(["bg", "en"])
   
   # 3) read
   result = reader.readtext("test.png")
   for bbox, text, conf in result:
       print(f"{text!r}  confidence={conf:.2f}  box={bbox}")

   Run it with python ocr_test.py. On the first run EasyOCR downloads the model weights into ~/.EasyOCR/model — you need internet and a little patience. Later runs start from the cache.

   The result is a list in which every row is a triple (box, text, confidence). We give no sample output because we have not run the code — you will see what was recognised and how sure the model is.

6. How to read the result

   • Box (bbox) — four corner points of the region, in pixels. Good for cropping, outlining or ordering.
   • Text — the recognised string.
   • Confidence (confidence) — a number between 0 and 1: how sure the model is. This is a model score, not a measured accuracy; compare it with your own samples before you trust it.

   The source can be a file path, an OpenCV (numpy) array, bytes or the URL of a raw image. If you only need the strings, without boxes:

   pythoncopy

   texts = reader.readtext("test.png", detail=0)
   print(" ".join(texts))
   
   # only the more confident results
   good = [t for (b, t, c) in reader.readtext("test.png") if c > 0.5]

   ✅

   One Reader per process

   The line easyocr.Reader([...]) loads the models into memory and takes time. Do it once and use the same reader for all images; do not create it inside a loop.

7. CPU instead of the GPU, working without internet

   By default gpu=True. If the GPU is not enough or PyTorch cannot see it, you switch to the CPU:

   pythoncopy

   reader = easyocr.Reader(["bg", "en"], gpu=False)

   We have not measured the speed of the two modes on your machine — run the same image in both modes and compare yourself. The library also has a command line (per the project's documentation): easyocr -l bg en -f test.png --detail=1 --gpu=True.

   Without internet: the weights can be downloaded beforehand from the project's model hub and placed by hand in ~/.EasyOCR/model. That way a second machine without network access can use them. The data you read stays with you either way — only the one-time download goes out.

   ⚠️

   The licence of the weights

   The library's licence is Apache-2.0. We have not checked the licence of the downloaded model weights separately — read the project's terms if you will use them in a product.

8. Strengths and limits

   The project describes EasyOCR as a general OCR that reads both text in scenes (signs, photos) and dense text in documents. What we know, and what we do not:

   Topic	What is known
   Interface	Two lines of Python; also a command line and a Dockerfile in the project's repository.
   Languages	Over 80, including Bulgarian (bg); languages combine only in allowed groups.
   Result	A box, text and confidence for every region.
   Handwriting	Not supported — it is on the project's roadmap as future work.
   Tables and page layout	The library returns regions with text; it does not reconstruct table structure. For complex documents you will need another tool.
   Maintenance	The latest release, 1.7.2, is from 24.09.2024; the repository automatically closes issues older than six months. Check compatibility with your PyTorch version.
   Accuracy and speed on Bulgarian	⚠️ Not measured by us. Collect 10–20 samples of your own and compare.

   If the text is on a clean, well-lit scan, the classic Tesseract on the CPU may be enough; for tables and complex pages look for a solution with structure — for example PP-StructureV3 from the PaddleOCR lesson.

9. Personal data — the minimum

   • Read documents with personal data only locally; do not upload them to cloud OCR services without a reason.
   • The results (text and boxes) may contain the same data — protect them like the originals.
   • Do not put real documents into public repositories or examples. Test with synthetic images, as in step 5.

04Check

• uname -m gives aarch64 and torch.cuda.is_available() gives True (or you know why you are on the CPU).
• import easyocr works inside the virtual environment.
• ocr_test.py returns triples (box, text, confidence) and what was recognised resembles what was written.
• With detail=0 you get only the strings.
• The gpu=False mode works too.
• You know what EasyOCR is not for (handwriting, tables).

Quiz

05What's next

06Sources

1. EasyOCR — repository and README 🌐 global — usage, readtext, detail=0, gpu=False, CRAFT and CRNN, handwriting as future work.
2. easyocr on PyPI — version 1.7.2 of 24.09.2024, Apache 2.0 licence.
3. Jaided AI: EasyOCR — table of supported languages (Bulgarian — bg).
4. EasyOCR: config.py — the Cyrillic language group, the model cache location.
5. EasyOCR: requirements.txt — dependencies.
6. PyTorch: CUDA 13.0 wheels · torchvision — aarch64 availability for CPython 3.12.
7. NVIDIA DGX Spark: hardware · ASUS Ascent GX10: tech specs — GB10 class, Arm, shared memory · known issues — "Memory-Usage: Not Supported" in nvidia-smi.