zchjkhzkhc/nte-courier-ocr

NTE & KAPY Courier Waybill & Customer Tag OCR (v1.0)

Specialized, ultra-lightweight OCR text recognition model tailored specifically for China - Vietnam cross-border courier waybills and warehouse customer tags.

Key Features & Benchmark

• Model Size: Only 5.12 MB (ONNX format).
• Latency: ~2.4 ms per text line (Intel CPU / Mobile NPU).
• Architecture: 5-block CNN feature extractor + 2-layer BiLSTM sequence model + CTC Greedy Decoder.
• Vocabulary: 116 pruned logistics characters (eliminating 6,500+ redundant characters for maximum speed and zero memory bloat).
• Zero Cloud Cost: Runs directly inside any mobile browser via WebAssembly / ONNX Runtime Web (no VPS, no GPU server needed).

Targeted Logistics Formats

1. Courier Waybills:
   • SF Express: SF162881597629, SF0107207480855...
   • Yuantong (YT): YT0215545579984, YT8804958066240...
   • J&T Express (JT): JT9957880404554, JT2229455338406...
   • Zhongtong (77/78): 775751644594606, 78796623403516...
2. Warehouse Customer Codes & Routing Marks:
   • CAPY233Duy, (CAPY233Duy)
   • NTE01HN-TONY, NTE555Hoc, NTE291 (D5)
   • 入仓号: 01HN::NTE 01HN
   • 01HN, 02SG, 01DN, HP01

Files in Repository

• nte_kapy_rec.onnx: Optimized ONNX model with dynamic width support.
• dict_nte.txt: 116-character vocabulary mapping (CTC blank = index 0).

WebAssembly / JavaScript Usage

import * as ort from 'onnxruntime-web';

// 1. Load ONNX Session in browser
const session = await ort.InferenceSession.create('https://huggingface.co/zchjkhzkhc/nte-courier-ocr/resolve/main/nte_kapy_rec.onnx');

// 2. Preprocess crop to 32px height
// 3. session.run({ input: tensor });