IlhaOCR: Edge-First Taiwanese Document & Healthcare OCR Engine

English | 繁體中文說明

Release Summary: IlhaOCR is an edge-optimized, in-process optical character recognition (OCR) engine engineered for Taiwanese Traditional Chinese identity documents (ROC National ID, National Health Insurance Card) and clinical healthcare paperwork. Distributed as lightweight, self-contained ONNX models, IlhaOCR runs 100% on-premise on resource-constrained edge kiosks, Raspberry Pi, and local clinical workstations without GPU acceleration or cloud API dependencies.

---

Key Highlights

• 🇹🇼 Taiwanese Typography & Terminology: Covers standard Traditional Chinese, rare personal names, official administrative divisions (縣市、鄉鎮市區、村里、鄰、街路巷弄號樓), clinical terms, and full alphanumeric character sets.
• 🔒 On-Premise Privacy: 100% local in-process inference designed to keep Protected Health Information (PHI) and identity data strictly within the local security perimeter.
• ⚡ Ultra-Low Edge Footprint:
  • 34.05 MB INT8 ONNX footprint for embedded hardware constraints.
  • < 30 ms single-vCPU thread inference latency.
  • Low peak RSS memory consumption (< 800 MB), suitable for fanless terminals and low-cost healthcare stations.
• 🏥 Healthcare & Identity Specialization:
  • National ID Card (國民身分證): Front & back extraction (Name, National ID checksum verification, ROC birthdate, issuance category 初發/換發/補發, address lines, and blank card serial number).
  • National Health Insurance Card (健保卡): 12-digit card number (0000 xxxx xxxx), National ID, and birthdate.
  • Clinical Documents: Triage records, clinic receipts, and laboratory forms.

---

Technical Specifications & Edge Benchmark

Metric / Specification	IlhaOCR (INT8 ONNX)	IlhaOCR (FP32 Baseline)	Embedded Budget
Architecture	MobileNetV3 + BiLSTM + CTC	MobileNetV3 + BiLSTM + CTC	Edge CRNN
Vocabulary	5,569 Tokens	5,569 Tokens	Traditional Chinese + Alphanumeric
Input Shape	[1, 1, 32, 320] (Grayscale)	[1, 1, 32, 320] (Grayscale)	Dynamic Width Supported
Model Size	34.05 MB (model_int8.onnx)	42.20 MB (model.onnx)	$\le 40$ MB
Single vCPU Latency	28.57 ms	41.20 ms	$\le 50$ ms
Peak Memory (RSS)	766 MB	820 MB	$\le 2$ GB
Runtime Dependency	onnxruntime (Zero PyTorch / CUDA)	onnxruntime	Pure CPU Inference

---

Quickstart

1. Installation

pip install onnxruntime numpy pillow huggingface_hub

2. Standalone Inference

import json
import numpy as np
import onnxruntime as ort
from PIL import Image
from huggingface_hub import hf_hub_download

# Download model and vocab from Hugging Face Hub
model_path = hf_hub_download(repo_id="weemed/IlhaOCR", filename="model_int8.onnx")
vocab_path = hf_hub_download(repo_id="weemed/IlhaOCR", filename="vocab.json")

with open(vocab_path, "r", encoding="utf-8") as f:
    vocab = json.load(f)["vocab"]
idx2char = {i: ch for i, ch in enumerate(vocab)}

# Initialize single-threaded edge session
opts = ort.SessionOptions()
opts.intra_op_num_threads = 1
opts.inter_op_num_threads = 1
session = ort.InferenceSession(model_path, opts, providers=["CPUExecutionProvider"])

def predict_crop(crop_img: Image.Image) -> str:
    # Resize and normalize to [1, 1, 32, 320]
    gray = np.array(crop_img.convert("L"))
    h, w = gray.shape
    scale = min(32 / max(h, 1), 320 / max(w, 1))
    new_w, new_h = max(1, int(w * scale)), max(1, int(h * scale))
    resized = np.array(Image.fromarray(gray).resize((new_w, new_h), Image.Resampling.BILINEAR))
    
    canvas = np.ones((32, 320), dtype=np.float32) * 255.0
    y_off = (32 - new_h) // 2
    canvas[y_off:y_off + new_h, 0:new_w] = resized
    norm = ((canvas / 255.0) - 0.5) / 0.5
    inp = norm[np.newaxis, np.newaxis, :, :].astype(np.float32)

    # Run inference
    logits = session.run(None, {"input": inp})[0]
    preds = np.argmax(logits[0], axis=-1)

    # CTC greedy decoding
    chars, prev = [], 0
    for p in preds:
        if p != 0 and p != prev:
            chars.append(idx2char.get(p, ""))
        prev = p
    return "".join(chars)

---

繁體中文說明

IlhaOCR 是專為台灣公衛醫療體系、基層診所長照機構與健康量測站（如 The Mirror）所研發之輕量化光學文字辨識（OCR）引擎。

核心特點

1. 地端隱私與個資保護（Zero PHI Leakage）：
   • 採用純 CPU-first ONNX 格式，推論全流程於本機記憶體完成，無需連網或上傳外部雲端，符合公衛與醫療隱私合規要求。
2. 極致邊緣端運算效能：
   • 經動態 INT8 量化後模型僅 34.05 MB。
   • 單核心 CPU（1 vCPU）辨識僅需約 28 毫秒。
   • 記憶體峰值控制於 766 MB，能無縫部署於樹莓派、嵌入式 Kiosk、無 GPU 之低成本邊緣主機。
3. 台灣官方證件精密適配：
   • 深入適配台灣身分證字體排版幾何特徵，精準捕捉發證字樣（初發、換發、補發）與背面雷射流水號（如 007871 6215）。
   • 內建台灣行政區劃辭彙（含鄉鎮市區、村里、鄰弄巷道門牌），提供高精確度的字元辨識。

---

醫療與身分資料安全邊界

• 預期用途（Intended Use）：身分證件登記輔助、臨床病歷攝取與健康量測站自動填單（Suggest-with-Review Assistant）。
• 人工覆核機制：辨識結果涉及身分證號校驗碼（Checksum）驗證與生日邏輯比對，建議於臨床登錄或正式掛號前保留操作人員確認界面。
• 資料保護：模型本體不儲存任何推論圖片或辨識文字，所有運算皆於臨時記憶體緩衝區執行完畢即釋放。

---

Citation & Governance

IlhaOCR is developed by WeeMed AI under the Apache 2.0 license.

@software{ilhaocr2026,
  author = {WeeMed AI Team},
  title = {IlhaOCR: Edge-First Taiwanese Document and Healthcare OCR Engine},
  url = {https://huggingface.co/weemed/IlhaOCR},
  year = {2026}
}