Date Expiration Model - YOLO + Custom OCR

This project converts annotation data into datasets suitable for training a date expiration detection model using YOLO for object detection and custom OCR for text recognition.

Dataset Overview

• Total Images: 11,858
• Training Set: 9,486 images (80%)
• Validation Set: 2,372 images (20%)
• Classes: prod (0), date (1), due (2), code (3)
• OCR Samples: 13,349 training + 3,325 validation text annotations

Directory Structure

custom-yolo/
├── convert.py              # ✅ Convert annotations to datasets
├── train_yolo.py           # 🚀 Train YOLOv11 model
├── train_ocr.py            # 🔤 Setup OCR training
├── inference.py            # 🔍 Run detection & recognition
├── run_all.py              # ⚡ Complete workflow
├── download_dataset.py     # 📥 Download dataset (gdown)
├── download_dataset_simple.py # 📥 Download dataset (requests)
├── test_portability.py     # 🧪 Test project portability
├── requirements.txt        # 📦 Dependencies
├── README.md               # 📖 Documentation
├── datasets/               # 📂 Original data (11,858 images)
│   ├── annotations.json    # Original annotations
│   └── images/            # Original images
├── yolo_dataset/           # 🟢 YOLO Training Data
│   ├── data.yaml          # YOLO configuration
│   ├── images/
│   │   ├── train/         # 9,486 training images
│   │   └── val/           # 2,372 validation images
│   └── labels/
│       ├── train/         # YOLO format labels
│       └── val/           # Validation labels
└── ocr_dataset/            # 🔵 OCR Training Data
    ├── annotations/
    │   ├── train_ocr.json  # 13,349 training samples
    │   └── val_ocr.json    # 3,325 validation samples
    └── images/
        ├── train/          # Training images (same as YOLO)
        └── val/            # Validation images

YOLO Dataset Format

Label Files (.txt)

Each label file corresponds to an image and contains lines in YOLO format:

class_id x_center y_center width height

Where:

• class_id: 0=prod, 1=date, 2=due, 3=code
• Coordinates are normalized to [0, 1]

Example:

0 0.1614785992217899 0.09536784741144415 0.15953307392996108 0.027247956403269755
1 0.377431906614786 0.09400544959128065 0.2178988326848249 0.02997275204359673

Data Configuration (data.yaml)

path: ./yolo_dataset  # Path to your yolo_dataset directory
train: images/train
val: images/val
nc: 4
names:
  - prod
  - date
  - due
  - code

OCR Dataset Format

Annotations (JSON)

Each OCR annotation contains:

{
  "image": "img_00001.jpg",
  "bbox": [138, 58, 250, 80],
  "text": "11.2024",
  "class": "date"
}

Where:

• image: Image filename
• bbox: [x1, y1, x2, y2] in absolute pixel coordinates
• text: The transcribed text
• class: The object class

Usage

🚀 Complete Workflow (Recommended)

# Run the complete pipeline
python run_all.py

# Or run individual steps:
python convert.py              # Convert data
python train_yolo.py --quick   # Train YOLOv11
python train_ocr.py            # Setup OCR
python inference.py --image test.jpg --visualize  # Test detection

1. Convert Data

python convert.py
# Creates yolo_dataset/ and ocr_dataset/

2. Train YOLOv11

# Quick training (recommended)
python train_yolo.py --quick

# Custom training
python train_yolo.py --model n --epochs 50 --batch 8

# Advanced training
python train_yolo.py --model m --imgsz 640 --epochs 100

3. Train OCR

python train_ocr.py
# Choose OCR method: EasyOCR, Tesseract, or Custom

4. Run Inference

# Single image
python inference.py --image test.jpg --visualize

# Batch processing
python inference.py --dir test_images/

# With trained model
python inference.py --model runs/train/weights/best.pt --image test.jpg

📊 Model Information

YOLOv11 Classes

• prod (0): Product labels
• date (1): Date text (with OCR)
• due (2): Due date text (with OCR)
• code (3): Barcodes/codes

Dataset Statistics

• Training images: 9,486
• Validation images: 2,372
• OCR training samples: 13,349
• OCR validation samples: 3,325

🔧 Requirements

pip install -r requirements.txt

Key dependencies:

• ultralytics>=8.3.0 - YOLOv11
• opencv-python - Image processing
• pytesseract - OCR (optional)
• easyocr - Advanced OCR (optional)

🎯 Quick Start

Option 1: Download Dataset Automatically

# Install dependencies
pip install -r requirements.txt

# Download and setup dataset
python download_dataset.py

# Run complete workflow
python run_all.py

# That's it! Your model is ready.

Option 2: Manual Dataset Setup

# Download dataset manually from:
# https://drive.google.com/file/d/1Ewnsqjm2SqdZvAN0EhVS6PulfoE59giX/view

# Extract to 'datasets' folder
# Then run:
python convert.py
python train_yolo.py --quick

Option 3: Simple Download (No gdown)

# If gdown fails, use simple version
python download_dataset_simple.py

📁 Dataset Download

Two download scripts are provided:

1. download_dataset.py - Uses gdown library (recommended)
2. download_dataset_simple.py - Uses requests only (fallback)

Both scripts will:

• ✅ Download dataset from Google Drive
• ✅ Extract ZIP file automatically
• ✅ Rename folder to datasets
• ✅ Clean up temporary files

🧪 Testing Portability

Before sharing, test that the project works on different systems:

python test_portability.py

This will verify:

• ✅ No hardcoded paths
• ✅ Works from different directories
• ✅ All scripts use dynamic paths
• ✅ Ready for sharing

📝 Notes

• YOLOv11: Latest version with improved accuracy and speed
• Portable: No hardcoded paths - works on any system
• Optimized: Reduced image sizes and batch sizes for edge deployment
• OCR Ready: Dataset prepared for text recognition training
• Minimal: Streamlined codebase with only essential files
• Shareable: Tested and verified to work when copied to different locations