YOLO Detection & OCR

Deep learning-based detection and text recognition for license plates.

Overview

The YOLO pipeline uses YOLOv11 for object detection and a custom YOLO OCR model for Arabic character recognition.

YOLO Pipeline

Detection Model

Architecture

YOLOv11 consists of three main components:

Backbone: CSPDarknet with C3k2 blocks for feature extraction
Neck: PANet (Path Aggregation Network) for multi-scale fusion
Head: Detection heads at P3, P4, P5 scales

Classes

Class	Description
`car`	Vehicle detection
`plate`	License plate detection

Usage

from ultralytics import YOLO

# Load model
model = YOLO("models/best.pt")

# Detect
results = model.predict(image, conf=0.25, iou=0.45)

# Process results
for box in results[0].boxes:
    cls_id = int(box.cls[0])
    cls_name = model.names[cls_id]
    x1, y1, x2, y2 = map(int, box.xyxy[0])

    if 'plate' in cls_name.lower():
        plate_crop = image[y1:y2, x1:x2]

OCR Model

Custom YOLO OCR

Instead of traditional OCR, we use a YOLO model trained to detect individual characters:

from src.pipeline.yolo_ocr import load_model, read_plate_text

# Load OCR model
ocr_model = load_model("models/yolo11m_car_plate_ocr.pt")

# Read plate
text, confidence = read_plate_text(plate_crop, model=ocr_model)
print(f"Plate: {text} ({confidence:.1%})")

Character Classes

The OCR model detects Arabic letters and numerals:

Type	Characters
Arabic Letters	ا ب ت ث ج ح خ د ذ ر ز س ش ص ض ط ظ ع غ ف ق ك ل م ن ه و ي
Numerals	0 1 2 3 4 5 6 7 8 9

Character Assembly

Characters are sorted left-to-right and assembled into the plate text:

def read_plate_text(plate_crop, model, conf=0.25):
    results = model.predict(plate_crop, conf=conf, verbose=False)

    detections = []
    for box in results[0].boxes:
        cls_id = int(box.cls[0])
        x_center = (box.xyxy[0][0] + box.xyxy[0][2]) / 2
        char = model.names[cls_id]
        detections.append((x_center, char))

    # Sort by x position (left to right)
    detections.sort(key=lambda x: x[0])

    # Assemble text
    text = ''.join([d[1] for d in detections])
    return text

Configuration

Detection Parameters

Parameter	Default	Description
`conf`	0.25	Confidence threshold
`iou`	0.45	NMS IoU threshold
`device`	auto	GPU/CPU selection

Model Files

models/
├── best.pt                    # Detection model
└── yolo11m_car_plate_ocr.pt   # OCR model

Training

Models were trained on Ultralytics HUB:

Detection: Custom Egyptian plate dataset
OCR: Character-level annotations on plate crops

See the Training Dashboard in the Streamlit app for metrics.

Comparison with EasyOCR

Feature	YOLO OCR	EasyOCR
Speed	Faster	Slower
Arabic Support	Trained	General
Accuracy	Higher for Egyptian	Lower for Arabic
Dependencies	Ultralytics	easyocr, torch