Automatic License Plate Recognition (ALPR) using Custom CNN & EasyOCR

A complete pipeline to detect and recognize vehicle license plates, designed as the core vision component for an automated toll collection system. This project features a custom 5-layer CNN for plate localization and leverages EasyOCR for character recognition.

---

► Key Features

• End-to-End System: Implements the full workflow from image input to recognized plate text.
• Custom Detector: A bespoke 5-layer CNN architecture for bounding box regression, trained from scratch in PyTorch.
• Data-Driven: Trained and validated on a cleaned dataset of ~8,000 images aggregated from multiple public sources.
• Reproducible: Notebooks for data preprocessing and model training are included.

---

🚀 Quick Look

Component	Technology / Method
Plate Detection	Custom 5-Layer CNN (PyTorch)
Character Recognition	EasyOCR
Dataset Size	~9,600 images (raw), ~8,000 (cleaned)
Input Resolution	416x416 (after aspect-ratio preserving resize)
Detection Loss	SmoothL1Loss
Training Techniques	Test-Time Augmentation (TTA), Dropout, Batch Normalization

---

📂 Repository Structure

ALPR-CNN-EasyOCR/
│
├── README.md
│
├── 1_Data_Exploration_and_Preprocessing.ipynb
├── 2_Training_and_Evaluation.ipynb
│
├── data/
│   ├── sample_images/
│   └── annotations/
│       ├── final_annotations.csv
│       └── final_annotations_preprocessed.csv
│
└── docs/
    ├── project_presentation.pdf
    └── mid_term_evaluation.pdf

---

🛠️ How It Works

1. Preprocessing: Input images are resized to 416x416 while preserving the original aspect ratio by adding padding. Bounding box coordinates are normalized.
2. Plate Detection: The custom CNN processes the image and predicts the four coordinates (xmin, ymin, xmax, ymax) of the license plate's bounding box.
3. Cropping: The predicted bounding box is used to crop the plate from the original, high-resolution image.
4. Character Recognition: The cropped plate image is passed to EasyOCR, which extracts the final license plate text.

---

📈 Training & Results

The model was trained using the Adam optimizer and a ReduceLROnPlateau learning rate scheduler. The notebooks provided in this repository detail the entire process:

• Data_Exploration_and_Preprocessing.ipynb: Covers how the raw dataset was loaded, cleaned, visualized, and prepared for training. This includes the logic for normalization and train/validation/test splits.
• Model_Training_and_Evaluation.ipynb: Contains the complete model definition, training loop, evaluation logic, and visualization of the loss curves.

For detailed metrics and performance analysis, please see the presentation and report in the /docs directory.

---

💡 Future Improvements

• Enhance OCR Accuracy: Fine-tune a dedicated OCR model on license plate character sets instead of using a general-purpose one.
• Upgrade the Backbone: Replace the custom CNN with a more powerful, pre-trained backbone (e.g., MobileNetV2, EfficientNet) for potentially higher detection accuracy.
• Real-time Video: Extend the pipeline to handle video streams by incorporating object tracking algorithms to maintain vehicle identity across frames.