PaddleOCR PP-OCRv5 models (Apache License 2.0, see LICENSE in this folder)
Copyright (c) PaddlePaddle Authors. https://github.com/PaddlePaddle/PaddleOCR

Original files: official ONNX inference models from the PaddleOCR distribution
  PP-OCRv5_mobile_det_onnx_infer.tar          (text detection)
  korean_PP-OCRv5_mobile_rec_onnx_infer.tar   (Korean text recognition)

Modified by this project (tools/shrink_ocr_models.mjs) — the files here are NOT the original files:
  PP-OCRv5_mobile_det.int8.onnx
    - Large convolution weights in the backbone stored as per-channel int8 + scale,
      restored to float inside the graph (Cast + Mul). Other weights unchanged.
  korean_PP-OCRv5_mobile_rec.int8.onnx
    - Convolution / MatMul weights stored as per-channel int8 + scale (restored with Cast + Mul).
    - Output layer reduced from 11947 to 2551 classes (blank, ASCII, Hangul compatibility jamo,
      the 2350 Hangul syllables of KS X 1001, characters used in game item names, space).
  korean_PP-OCRv5_mobile_rec.dict.json
    - Character list from the original inference.yml (character_dict), reduced to the same characters.