@llein/anpr-plate-cleaner

Production-grade OCR / ANPR license plate text cleaner, character confusion disambiguation (0 vs O, 1 vs I, 8 vs B), 2-line combiner, and whitelist fuzzy matcher.

⚡ Why This Package?

In real-world parking barrier systems and traffic cameras (ANPR / ALPR), raw OCR output from models like YOLO, PaddleOCR, or EasyOCR is often corrupted by:

• Night infrared reflections, dirt, scratches, and screw heads on the license plate.
• OCR character confusions:
  • Letters read where digits are required (e.g., S instead of 5, I / L instead of 1, B instead of 8, O / D instead of 0).
  • Digits read where letters are required (e.g., 8 instead of B, 0 instead of D/O, 1 instead of I).
• Two-line plates (top line: province & series, bottom line: sequence numbers).

@llein/anpr-plate-cleaner solves this using position-aware syntactic disambiguation, 2-line merging, and confusion-discounted fuzzy matching for resident whitelists.

🚀 Features

• Position-Aware Error Correction:
  • Automatically fixes province digits, series letters, and serial numbers according to vehicle registration standards (Circular 24/2023/TT-BCA).
• Two-Line Plate Stitching:
  • Combines multi-line OCR detections (e.g. 59-P1 on line 1 and 123.45 on line 2 -> 59P112345).
• Vehicle Type Classification:
  • Classifies into car, motorbike, electric_motorbike, military, and diplomatic.
• Confusion-Weighted Fuzzy Matching:
  • Levenshtein distance with fractional penalties for known OCR confusions (e.g. mistaking 8 for B is penalized with only 0.25 instead of 1.0).
  • Perfect for matching parking gate scans against database whitelist records.
• Zero External Dependencies: Pure TypeScript, minimal footprint (< 4KB).

📦 Installation

# Using npm
npm install @llein/anpr-plate-cleaner

# Using pnpm
pnpm add @llein/anpr-plate-cleaner

# Using yarn
yarn add @llein/anpr-plate-cleaner

🛠️ Usage Examples

1. Clean Corrupted OCR Output

import { cleanVietnamPlate, cleanANPRText } from '@llein/anpr-plate-cleaner';

// OCR misread 'S' as 5, 'I' as 1, 'B' as 8
const result = cleanVietnamPlate('SIK-999.9B');

console.log(result.compact);   // "51K99998"
console.log(result.formatted); // "51K-999.98"
console.log(result.vehicleType); // "car"
console.log(result.isValid);   // true

2. Multi-line Motorbike Plate

// Bounding boxes read top & bottom lines
const rawOCR = "59-P1\n123.45";

const result = cleanVietnamPlate(rawOCR);
console.log(result.compact);   // "59P112345"
console.log(result.formatted); // "59-P1 123.45"
console.log(result.vehicleType); // "motorbike"

3. Match Plate Against Parking Whitelist (Fuzzy Search)

import { findBestPlateMatch } from '@llein/anpr-plate-cleaner';

const whitelist = [
  '51K99999',
  '29A12345',
  '59P167890'
];

// OCR reading has a glitch: 'O' instead of '0'
const scanned = '59P16789O';

const match = findBestPlateMatch(scanned, whitelist, 0.85);

if (match) {
  console.log(`Open Barrier! Matched: ${match.match} (Confidence: ${match.score * 100}%)`);
}

📖 API Reference

cleanVietnamPlate(raw: string): PlateCleanResult

Full syntactic analysis returning { raw, compact, formatted, vehicleType, isValid, score }.

cleanANPRText(raw: string, options?: CleanOptions): string

Returns cleaned string directly (compact or formatted).

combinePlateLines(topLine: string, bottomLine: string): string

Joins 2-line OCR boxes into a unified text string.

calculatePlateSimilarity(plateA: string, plateB: string): number

Returns a similarity score between 0.0 and 1.0 utilizing OCR confusion matrices.

findBestPlateMatch(query: string, candidates: string[], threshold?: number)

Searches a candidate array for the closest matching registered plate.

🧪 Testing

npm test

📄 License

MIT © Llein-Dev