YOLO OCR

Custom YOLO model trained to detect and recognize Arabic characters on Egyptian license plates.

Overview

Instead of traditional OCR engines, we use a YOLO object detection model where each Arabic character and numeral is a separate class. This approach is more robust for the specific format of Egyptian plates.

Architecture

The YOLO OCR model detects individual characters as objects:

Plate Crop → YOLO Inference → Character Boxes → Sort Left-to-Right → Assemble Text

Character Classes

Arabic Letters (17 classes)

Character	Name	Character	Name
ا	Alef	ر	Ra
ب	Ba	س	Seen
ج	Jeem	ص	Sad
د	Dal	ط	Ta
ع	Ain	ف	Fa
ق	Qaf	ل	Lam
م	Meem	ن	Noon
ه	Ha	و	Waw
ي	Ya

Numerals (10 classes)

| 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |

Usage

from src.pipeline.yolo_ocr import load_model, read_plate_text

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

# Read plate text
plate_crop = cv2.imread("plate.jpg")
text, confidence = read_plate_text(plate_crop, model=model)

print(f"Plate: {text} ({confidence:.1%})")

Character Assembly

Characters are sorted by their x-coordinate (left to right) and concatenated:

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 = CLASS_MAP.get(model.names[cls_id], model.names[cls_id])
        detections.append((x_center, char, float(box.conf[0])))

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

    # Build text
    text = ''.join([d[1] for d in detections])
    avg_conf = sum(d[2] for d in detections) / len(detections) if detections else 0

    return text, avg_conf * 100

Class Mapping

The model outputs class names that need to be mapped to actual characters:

CLASS_MAP = {
    "alef": "ا", "ba": "ب", "jeem": "ج", "dal": "د",
    "ra": "ر", "seen": "س", "sad": "ص", "ta": "ط",
    "ain": "ع", "fa": "ف", "qaf": "ق", "lam": "ل",
    "meem": "م", "noon": "ن", "ha": "ه", "waw": "و",
    "ya": "ي",
    "zero": "0", "one": "1", "two": "2", "three": "3",
    "four": "4", "five": "5", "six": "6", "seven": "7",
    "eight": "8", "nine": "9"
}

Configuration

Parameter	Default	Description
`conf`	0.25	Detection confidence threshold
`iou`	0.45	NMS IoU threshold
`verbose`	False	Print detection output

Model File

models/
└── yolo11m_car_plate_ocr.pt   # YOLO OCR weights (~40MB)

Comparison with EasyOCR

Feature	YOLO OCR	EasyOCR
Speed	⚡ Faster	Slower
Arabic Accuracy	✅ Higher (trained)	Lower (general)
Setup	Ultralytics	easyocr package
Customization	✅ Trainable	Limited

Training

The model was trained on Ultralytics HUB with:

Dataset: Custom Egyptian plate character annotations
Base Model: YOLOv11m
Classes: 27 (17 Arabic letters + 10 numerals)
Augmentation: Rotation, blur, brightness variation