Face Classification using MobileNet & Transfer Learning

A face classification system built with MobileNet (transfer learning), face_recognition for preprocessing, and a Streamlit web interface for real-time inference. The pipeline handles everything from raw image preprocessing and augmentation through model training to deployment via Docker.

How It Works

The project has two main components:

Training pipeline (face-recog-mobile.py) takes a folder of labeled face images, resizes them to 224×224, validates that each image contains a detectable face (discarding those that don't), splits the data into train/validation/test sets (70/20/10), applies image augmentation (50 augmented copies per training image), and fine-tunes a MobileNet model with a custom classification head. The trained model is saved as face-mobile.h5.

Inference app (app.py) loads the trained model into a Streamlit web app where users can upload an image and get a predicted class label with confidence score.

Project Structure

├── data/                    # Raw training images (one subfolder per person/class)
│   ├── person_1/
│   ├── person_2/
│   └── ...
├── processed/               # Auto-generated train/val/test splits + augmented data
├── no-face/                 # Images where no face was detected (auto-moved)
├── shells/                  # Shell scripts
├── face-recog-mobile.py     # Training pipeline
├── app.py                   # Streamlit inference app
├── face-mobile.h5           # Trained model weights (~16 MB)
├── accuracy_plot.png         # Training/validation accuracy over epochs
├── confusion_matrix.png     # Test set confusion matrix
├── Dockerfile               # Container config for deployment
├── requirements.txt         # Python dependencies
└── README.md

Setup

Prerequisites

• Python 3.9+
• dlib (required by face_recognition — may need CMake installed)

Installation

git clone <repo-url>
cd face-recog-opencv

pip install -r requirements.txt

Prepare Training Data

Organize your images under the data/ directory with one subfolder per class:

data/
├── alice/
│   ├── img1.jpg
│   ├── img2.jpg
│   └── ...
├── bob/
│   ├── img1.jpg
│   └── ...

Note: The final Dense layer in face-recog-mobile.py is set to 6 output units. Update Dense(6, ...) to match your actual number of classes before training.

Usage

Train the Model

python face-recog-mobile.py

This will:

1. Resize all images to 224×224
2. Detect and filter out images with no faces (moved to no-face/)
3. Split data into train, validation, and test sets
4. Generate 50 augmented images per training sample
5. Fine-tune MobileNet for 10 epochs
6. Save the model to face-mobile.h5
7. Output accuracy_plot.png and confusion_matrix.png

Run the Web App

streamlit run app.py

Open http://localhost:8501 in your browser, upload a face image, and get the classification result.

Run with Docker

docker build -t face-classifier .
docker run -p 8501:8501 face-classifier

Model Architecture

The model uses MobileNet pretrained on ImageNet as a frozen feature extractor, with a custom head:

MobileNet (frozen) → GlobalAveragePooling2D → Dense(256, ReLU) → Dense(num_classes, Softmax)

Training uses Adam optimizer with categorical crossentropy loss. Image augmentation includes rotation, shifts, shear, zoom, and horizontal flips.

Key Dependencies

Package	Purpose
tensorflow/keras	Model training and inference
face_recognition	Face detection and validation
streamlit	Web interface
scikit-learn	Data splitting and evaluation
matplotlib	Accuracy plots
seaborn	Confusion matrix visualization
Pillow	Image loading and resizing
opencv-python	Image processing utilities

License

See repository for license details.