Celeb Look-Alike (End-to-End Deep Learning)

A Cyberpunk-themed AI web application that identifies your celebrity twin using Face Embeddings, Transfer Learning, and Vector Similarity Search.

---

📸 Project Demo

---

🚀 Key Features

• End-to-End Pipeline: Covers the complete lifecycle from raw image ingestion to production inference.
• Cyberpunk UI: Neon-styled frontend with real-time Webcam Integration using HTML5 + JavaScript.
• Robust Face Extraction: Automatic face detection, alignment, and cropping using MediaPipe.
• Deep Learning Core: Fine-tuned ResNet50 achieving ~83.4% validation accuracy on a custom dataset.
• Vector Similarity Search: One-shot learning using 128-D L2-normalized embeddings and Cosine Similarity.

---

🏗️ System Architecture

The system follows a modular Component–Pipeline Architecture.

graph TD
    classDef user fill:#ff9,stroke:#333,stroke-width:2px,color:black;
    classDef frontend fill:#00f7ff,stroke:#000,stroke-width:2px,color:black;
    classDef backend fill:#39ff14,stroke:#000,stroke-width:2px,color:black;
    classDef ai fill:#ff00ff,stroke:#000,stroke-width:2px,color:white;
    classDef data fill:#ffa500,stroke:#000,stroke-width:2px,color:black;

    User(["👤 User"]) --> UI["💻 Web Interface<br/>(HTML / JS / Cyberpunk CSS)"]
    UI --> Flask["🔥 Flask Server<br/>(app.py)"]

    subgraph Inference ["⚡ Inference Pipeline"]
        Flask --> MP["📷 Face Extractor<br/>(MediaPipe)"]
        MP --> RN["🧠 Embedding Model<br/>(ResNet50)"]
        RN --> Vec["128-D Vector<br/>L2 Normalized"]
    end

    subgraph Search ["🔍 Similarity Search"]
        Vec --> Match["🧮 Cosine Similarity"]
        DB[("🗄️ Stored Embeddings<br/>Pickle")] --> Match
    end

    Match --> Output["✅ Match Result"]
    Output --> Flask

    subgraph Training ["🏗️ Offline Training"]
        Raw["📂 Dataset"] --> Aug["🔄 Augmentation"]
        Aug --> Trainer["🏋️ Trainer"]
        Trainer --> RN
        Trainer --> DB
    end

    class User user;
    class UI frontend;
    class Flask,MP,Match,Output backend;
    class RN,Vec,Trainer,Aug ai;
    class DB,Raw data;

Loading

---

🛠️ Tech Stack

Domain	Tools
Language	Python 3.11
Deep Learning	TensorFlow 2.15, Keras (Transfer Learning)
Computer Vision	MediaPipe 0.10.9, OpenCV, Pillow
Backend	Flask 3.1.2, Gunicorn
Frontend	HTML5, CSS3 (Neon Theme), JavaScript (Webcam API)
Data	NumPy, Pickle, Scikit-learn

---

⚙️ Installation & Setup

1️⃣ Clone the Repository

git clone https://github.com/your-username/celeb-look-alike.git
cd celeb-look-alike

2️⃣ Create Environment

conda create -n celeb python=3.11 -y
conda activate celeb

3️⃣ Install Dependencies

pip install -r requirements.txt

4️⃣ Run the Application

python app.py

Access the app at http://127.0.0.1:5000 and enable webcam capture or upload an image.

---

🧠 Model Training Details

Phase 1: Warm-Up Training

• Base Model: ResNet50 (ImageNet weights, frozen)
• Outcome: ~40% accuracy (generic feature learning)

Phase 2: Fine-Tuning

• Unfrozen top ResNet layers
• Final Validation Accuracy: ~83.4%

---

📂 Project Structure

├── data                   # Cropped faces dataset used during embedding genration
├── artifacts/             # Models (.h5) and embeddings.pkl
├── notebooks/             # Experiments & EDA
├── src/
│   ├── components/        # Data ingestion, trainer, generators
│   ├── pipelines/         # Training & prediction pipelines
│   ├── exception.py       # Custom exceptions
│   ├── logger.py          # Logging
│   └── utils.py           # Face extraction & helpers
├── static/
│   ├── css/               # Cyberpunk styling
│   └── js/                # Webcam logic
|   └── uploads/           # Temporary hold the image
├── templates/             # HTML templates
├── app.py                 # Flask entry point
└── requirements.txt       # Dependencies

---

🤝 Credits

• Developer: Aryan Gairola
• Dataset: kaggle link : "vishesh1412/celebrity-face-image-dataset"