anaksor-ocr

anaksor-ocr

Reads Khmer and English text from documents, receipts and photos of signs, using a detector for text lines and a recogniser for each line. Suits developers who need Khmer OCR in Python.

Made by Sokheng on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: anaksor-ocr holds the weights for Khmer OCR, with English and Latin text supported too. Two models work together. A detector finds the lines of text in an image, and a recogniser reads each cropped line. They power the anaksor-ocr Python package.

Who it suits: Developers who need to turn Khmer documents, receipts and sign photos into text, under the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Administration | Records teams | Digitize Khmer letters and forms | Turn a scanned letter into text
Retail and finance | Back-office teams | Read receipts that mix Khmer and Latin text | Pull the text from a receipt photo
Software | App developers | Add Khmer OCR to an app | Read a sign photo in a mobile app
Research | NLP researchers | Study Khmer text detection and recognition | Compare model versions on a test set

Skip it if your text is in other scripts, since the model covers Khmer and English.

Where to use it: It runs in Python through the anaksor-ocr package, or from its command line.

• Python on Windows, macOS or Linux
• ONNX Runtime, which needs no PyTorch
• The PyTorch backend, with CUDA or Apple MPS support
• Offline machines, with weights placed in a local folder

Use case | Who | What they get back
Document OCR | Back-office teams | The text of a letter or form
Receipt reading | Finance teams | The text lines found on a receipt
Sign reading | App developers | Text recognised in a photo

How to set it up: You need Python and the package.

• Python with pip
• The anaksor-ocr package, with the torch extra if you want the PyTorch backend
• A few hundred MB at most of weights, which the package fetches for you

1. Install the anaksor-ocr package with pip.
2. Optionally add the torch extra for the PyTorch backend.
3. Run it once so the package downloads and caches the model files.

pip install anaksor-ocr
pip install "anaksor-ocr[torch]"

You can also download the repository by hand, and pin a release by its tag.

pip install -U huggingface_hub
hf download Sokheng/anaksor-ocr --local-dir weights/huggingface
hf download Sokheng/anaksor-ocr --revision rec-v4-det-dbpp-v1 --local-dir weights/huggingface

How to use it: Create the OCR object, pass it an image path, and read the text from the result.

import anaksor

ocr    = anaksor.OCR()
result = ocr("letter.jpg")
print(result["text"])

The command line does the same. Environment variables pin a release or use local weights with no network.

ANAKSOR_HF_REVISION=rec-v4-det-v3 anaksor letter.jpg
anaksor --fetch-models
ANAKSOR_WEIGHTS_DIR=./weights anaksor letter.jpg

Tool | What it does
anaksor.OCR | Runs detection and recognition on an image
anaksor | Command line that reads an image file
anaksor --fetch-models | Downloads the model files ahead of time
ANAKSOR_HF_REVISION | Pins a tagged release of the weights
ANAKSOR_WEIGHTS_DIR | Uses local weights and no network

Example: You pass a photo of a typed Khmer letter to the OCR object. The result comes back with a text field, and you print it to read the letter's text.

For newcomers and experts: If you are new to OCR, install the package, run it on one clear photo and print the text. Fetching the models first is a good idea if your network is slow.

If you know the field, recognition is an SVTRv2-B encoder with a CTC head, and detection is a mobile DBNet++ with a RepViT backbone and an Adaptive Scale Fusion block. The package hash-checks each download and caches it. The card adds that loading these file types does not execute code. Load a .pth checkpoint with the .yml config from the same revision.

Load an image
Detector finds the text lines
Crop each line
Recogniser reads each crop
Return the text

Good to know: The card reports recognition v4 at 86.7% line accuracy on real validation lines, against 84.8% for the previous version. Gains are largest on Latin text, while Khmer-only lines are about one point behind v3.

Title: Line accuracy on real validation lines
v3 all lines: 84.8%
v4 all lines: 86.7%
v3 Latin lines: 66.9%
v4 Latin lines: 76.4%

• License: Apache 2.0
• Cost to run: none named, apart from your own hardware
• Data it reaches: the images you give it, plus model downloads and a local cache under your home folder
• Not for: scripts other than Khmer and Latin, and reading text without checking the output

On Hugging Face, 14 November 2026

• Task: image to text
• Library: pytorch
• Formats: Safetensors
• Downloads in the last 30 days: 61
• Likes: 4
• License: apache-2.0
• Languages: km, en
• Tags: pytorch, onnx, khmer, text-detection, text-recognition, svtrv2, dbnet, dbnetpp, safetensors, image-to-text

Links

• Model: https://huggingface.co/Sokheng/anaksor-ocr
• Files: https://huggingface.co/Sokheng/anaksor-ocr/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Sokheng/anaksor-ocr

Then follow the usage in its model card. Nothing in this kit runs by itself.