mer 1.2.1

Khmer OCR

Mer (មើល)

Mer (មើល) is a lightweight bilingual (Khmer/English) OCR recognizer built around my custom CNN-Transformer network exported to ONNX.
This repository now focuses exclusively on line-level OCR — the Surya-powered layout, table, and LaTeX helpers have been removed, so the package is small and easy to embed anywhere you just need text from a single line image.

Installation

pip install mer

Getting started

from mer import Mer, postprocess_text

ocr = Mer()
ocr.load()  # optional; kept for backwards compatibility

text = ocr.recognize_line("samples/sample_1.png")
print("Line text:", text)

# predict() is an alias for recognize_line()
json_result = ocr.predict("samples/sample_1.png", json_result=True)
print(json_result["text"])

# Optional extra cleanup
print(postprocess_text("ទៀតផង ។"))  # -> "ទៀតផង។"

Configuration options

All options control the ONNX Runtime predictor:

• device: "cpu", "cuda", or specific device strings. Defaults to "cuda" with automatic CPU fallback.
• providers: optional explicit ONNX Runtime provider list. When omitted, providers are derived from device.
• model_path: point to a directory containing khmer_ocr.onnx and config.json to skip Hugging Face downloads.
• cache_dir / repo_id: control where artifacts are downloaded from Hugging Face Hub (metythorn/ocr-stn-cnn-transformer-base by default).
• max_length: override the configured maximum decoding length.
• postprocess: disable built-in whitespace cleanup if you prefer the raw model output.
• json_result: default return type for predict(). When True, predict() returns {"text": ...}; otherwise it returns a raw string. You can always override this per-call.

Using local model files

If you already have the ONNX weights and config on disk, point Mer at the folder to skip any Hugging Face download:

from mer import Mer

ocr = Mer(model_path="/path/to/local/model_dir", device="cpu")
ocr.load()
print(ocr.recognize_line("line.png"))

Post-processing helper

postprocess_text is exposed as a standalone helper so you can reuse the same Khmer punctuation cleanup on your own strings:

from mer import postprocess_text

assert postprocess_text(" ទៀត\tផង  ។ ") == "ទៀត ផង។"

Sample data

The samples/ directory contains a few PNGs you can use for quick manual testing. They are untouched and meant purely for experimentation with the line recognizer.