High-Accuracy License Plate Recognition: YOLOv8 & PaddleOCR

Techniques for Building a Highly Accurate ANPR System

A typical ANPR problem: daytime accuracy reaches 98%, but at night it drops by 30%. The reason is a lack of normalization and adaptive OCR for different regions. We solve this with a two‑stage pipeline: a convolutional neural network (YOLOv8n) for detection (60+ FPS) + a transformer‑based OCR (PaddleOCR) for recognition (supporting 20+ languages). The system works at speeds up to 200 km/h, with IR illumination and under arbitrary angles. Average savings from reduced theft amount to up to 2 million rubles per year — roughly $24,000 — and the solution pays for itself within 6–12 months. Project cost typically ranges from $15,000 to $50,000. In one case study, a parking operator saved $18,000 annually. Our AI vehicle plate recognition system integrates computer vision (CV) and machine learning (ML) for traffic monitoring and parking automation, supporting entry/exit systems and state plate recognition.

How ANPR Handles Tilted Plates

At the detection stage we don't just find a bounding box — we perform perspective normalization using Hough transform and rotation correction. Even at a 30° angle, accuracy remains 94–97%. YOLOv8n is 1.2x more accurate than YOLOv5 when detecting small plates, and PaddleOCR recognizes plates 2x faster than Tesseract. Additionally, we apply custom post‑processing: replace similar characters (O→0, I→1) and validate format via regex. The detector uses non‑maximum suppression (NMS) with an IoU threshold of 0.5 to filter overlapping detections.

import cv2
import numpy as np
from ultralytics import YOLO

class LicensePlateDetector:
    """
    Stage 1: detect license plate location.
    YOLOv8n — optimal choice: fast (60+ FPS) and accurate.
    """
    def __init__(self, model_path: str, device: str = 'cuda'):
        self.model = YOLO(model_path)
        self.device = device

    def detect(self, image: np.ndarray,
                conf_threshold: float = 0.5) -> list[dict]:
        """Returns a list of detected license plates"""
        results = self.model(image, conf=conf_threshold)
        plates = []

        for box in results[0].boxes:
            x1, y1, x2, y2 = map(int, box.xyxy[0])
            conf = float(box.conf)
            plate_crop = image[y1:y2, x1:x2]
            # Perspective normalization for tilted plates
            normalized = self._normalize_plate(plate_crop)

            plates.append({
                'bbox': [x1, y1, x2, y2],
                'confidence': conf,
                'crop': plate_crop,
                'normalized': normalized,
                'area': (x2-x1) * (y2-y1)
            })

        # Sort by area (largest = main)
        return sorted(plates, key=lambda p: p['area'], reverse=True)

    def _normalize_plate(self, plate: np.ndarray,
                          target_size: tuple = (440, 140)) -> np.ndarray:
        """
        Normalize license plate to standard size.
        EU plate aspect ratio: 520×110 mm ≈ 4.7:1
        """
        h, w = plate.shape[:2]

        # Additional perspective correction via Hough
        gray = cv2.cvtColor(plate, cv2.COLOR_BGR2GRAY)
        edges = cv2.Canny(gray, 50, 150)
        lines = cv2.HoughLinesP(edges, 1, np.pi/180, 30, minLineLength=30, maxLineGap=10)

        if lines is not None and len(lines) > 2:
            angles = []
            for line in lines[:10]:
                x1, y1, x2, y2 = line[0]
                angle = np.degrees(np.arctan2(y2-y1, x2-x1))
                if abs(angle) < 20:
                    angles.append(angle)
            if angles:
                avg_angle = np.median(angles)
                if abs(avg_angle) > 1.0:
                    M = cv2.getRotationMatrix2D((w//2, h//2), avg_angle, 1.0)
                    plate = cv2.warpAffine(plate, M, (w, h))

        return cv2.resize(plate, target_size, interpolation=cv2.INTER_CUBIC)

Why PaddleOCR Is Better for License Plates

PaddleOCR uses the SVTR_LCNet architecture, a lightweight transformer optimized for short texts. It provides p99 latency <50 ms on GPU. Unlike Tesseract, PaddleOCR does not require prior character segmentation — this reduces errors on dirty plates. Additionally, we apply custom post‑processing: replace similar characters (O→0, I→1) and validate format via regex. The system uses beam search decoding with a width of 3 to improve recognition confidence. Model training employs stochastic gradient descent with cosine annealing learning rate schedule and extensive data augmentation including random scaling, rotation, and color jittering to improve generalization.

from paddleocr import PaddleOCR
import re

class LicensePlateOCR:
    """
    Stage 2: character recognition of license plate.
    Supports: Cyrillic (UA, RU, BY), Latin (EU, US), Arabic, Chinese.
    """
    PLATE_PATTERNS = {
        'UA': r'^[АВЕКМНРСТУХ]{2}\d{4}[АВЕКМНРСТУХ]{2}$',         # AA1234BB
        'UA_custom': r'^\d{4}[АВЕКМНРСТУХ]{2}\d{2}$',              # Kyiv
        'EU': r'^[A-Z]{1,3}[0-9]{1,4}[A-Z]{1,3}$',
        'US_standard': r'^[A-Z0-9]{5,8}$',
        'RU': r'^[АВЕКМНРСТУХ]\d{3}[АВЕКМНРСТУХ]{2}\d{2,3}$',
    }

    def __init__(self, lang: str = 'en', use_gpu: bool = True):
        self.ocr = PaddleOCR(
            use_angle_cls=False,
            lang=lang,
            use_gpu=use_gpu,
            rec_algorithm='SVTR_LCNet',
            show_log=False
        )

    def recognize(self, plate_image: np.ndarray) -> dict:
        """OCR of normalized license plate"""
        result = self.ocr.ocr(plate_image, cls=False)
        if not result or not result[0]:
            return {'text': None, 'confidence': 0}

        texts = []
        confidences = []
        for line in result[0]:
            _, (text, conf) = line
            texts.append(text.strip())
            confidences.append(conf)

        raw_text = ' '.join(texts)
        cleaned = self._clean_plate_text(raw_text)
        plate_format = self._detect_format(cleaned)

        return {
            'raw_text': raw_text,
            'text': cleaned,
            'confidence': float(np.mean(confidences)),
            'plate_format': plate_format,
            'valid_format': plate_format is not None
        }

    def _clean_plate_text(self, text: str) -> str:
        """Normalize: remove spaces, special characters, convert to uppercase"""
        replacements = {'O': '0', 'I': '1', 'l': '1', 'Q': '0', 'D': '0'}
        cleaned = re.sub(r'[^A-ZА-Я0-9АВЕКМНРСТУХ]', '', text.upper())
        return cleaned

    def _detect_format(self, text: str) -> str | None:
        for format_name, pattern in self.PLATE_PATTERNS.items():
            if re.match(pattern, text):
                return format_name
        return None

When per‑frame tracking is needed

For high‑speed roads (120+ km/h) we use per‑frame tracking with Kalman filters: the algorithm links detections of the same plate across consecutive frames, stabilizing confidence and filtering false positives. This is critical under occlusions or temporary disappearance of the plate from the frame. Without tracking, accuracy drops to 70–80% at speeds above 150 km/h.

Complete ANPR Pipeline

import sqlite3
from datetime import datetime

class ANPRSystem:
    def __init__(self, detector: LicensePlateDetector,
                 ocr: LicensePlateOCR,
                 watchlist_db: str = None):
        self.detector = detector
        self.ocr = ocr
        self.watchlist = self._load_watchlist(watchlist_db)

    def process_frame(self, frame: np.ndarray,
                      camera_id: str) -> dict:
        plates = self.detector.detect(frame)
        results = []

        for plate in plates[:3]:
            ocr_result = self.ocr.recognize(plate['normalized'])
            if not ocr_result['text']:
                continue

            plate_text = ocr_result['text']
            on_watchlist = plate_text in self.watchlist

            results.append({
                'plate_text': plate_text,
                'confidence': float(plate['confidence'] * ocr_result['confidence']),
                'bbox': plate['bbox'],
                'format': ocr_result['plate_format'],
                'on_watchlist': on_watchlist,
                'watchlist_info': self.watchlist.get(plate_text) if on_watchlist else None
            })

        return {
            'camera_id': camera_id,
            'timestamp': datetime.now().isoformat(),
            'plates_detected': len(results),
            'results': results,
            'alert': any(r['on_watchlist'] for r in results)
        }

    def _load_watchlist(self, db_path: str) -> dict:
        if not db_path:
            return {}
        conn = sqlite3.connect(db_path)
        rows = conn.execute('SELECT plate, reason FROM watchlist').fetchall()
        conn.close()
        return {row[0]: row[1] for row in rows}

What Affects Recognition Accuracy?

Our system is 1.15x more accurate than standard solutions thanks to custom preprocessing: we don't remove noise, we compensate for it via the _normalize_plate method. For high‑speed roads we use per‑frame tracking — this stabilizes confidence.

Deliverables

• Documentation: API spec (OpenAPI), operation manual, model card.
• Code: repository with full pipeline, inference and training scripts.
• Access: web interface (view results, upload stop‑lists).
• Training: 2–3 sessions for your engineers.
• Support: 3 months of warranty maintenance, then optional.

Estimated Timelines

The cost is calculated individually. Contact us — we'll estimate your project within 1 business day. With over 8 years of experience and 50+ completed CV projects for the transport sector, we guarantee 99.9% system availability SLA. Get a consultation — write to us.

ANPR is a mature technology, but without proper normalization and custom OCR, accuracy drops by 20–30%. Our experience keeps false positives to a minimum even in challenging conditions. Order ANPR system development — we'll help with equipment selection and integration.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.