😷 Face Mask Detector

A two-stage deep learning pipeline that detects faces in any photo and classifies each one as with mask or without mask — with colored bounding boxes and confidence scores drawn on the output image.

---

📊 Performance

Metric	Score
Validation Accuracy	~98–99%
Validation Loss	< 0.05
Training Time	~5–8 min (T4 GPU)
Face Detection	Multi-face, any angle

---

🔁 Pipeline Architecture

        [ Input Image ]
               │
  ┌────────────▼────────────┐
  │   MediaPipe BlazeFace   │  ← handles partial faces,
  │   Face Detection        │    angles, group photos
  └────────────┬────────────┘
               │  (one padded crop per face)
  ┌────────────▼────────────┐
  │   Preprocessing         │  ← preprocess_input is embedded
  │   (inside the model)    │    inside the model graph,
  │   Resize → 224×224      │    not applied manually
  └────────────┬────────────┘
               │
  ┌────────────▼────────────┐     ┌─────────────────────────┐
  │   MobileNetV2 Backbone  │  +  │   Classification Head   │
  │   Phase 1: fully frozen │     │   GlobalAveragePooling  │
  │   Phase 2: top 30 layers│     │   Dense(128) + Dropout  │
  │           unfrozen      │     │   Dense(1) → Sigmoid    │
  └────────────┬────────────┘     └─────────────────────────┘
               │
      ┌────────┴────────┐
      ▼                 ▼
  ✅ With mask      ❌ No mask
  + confidence      + confidence

---

✨ Key Features

• Transfer learning — MobileNetV2 pretrained on ImageNet provides strong visual features without training from scratch.
• Two-phase fine-tuning — Phase 1 trains only the custom head with the backbone frozen. Phase 2 unfreezes the top 30 MobileNetV2 layers and continues at a low learning rate (1e-5) to avoid destroying pretrained weights.
• MediaPipe face detection — BlazeFace (MediaPipe's engine) replaces Haar Cascades. It handles angled faces, partial occlusions, and multiple faces in a single photo far more reliably.
• Integrated preprocessing — preprocess_input is embedded inside the Keras model graph rather than applied in the inference function. This eliminates a common deployment bug where the training and inference scaling steps get out of sync.
• Fast data pipeline — tf.data.image_dataset_from_directory with parallel mapping and prefetch(AUTOTUNE) replaces ImageDataGenerator, removing the CPU bottleneck during training.
• Padded face crops — Each detected face is padded by 15% before classification, giving the model context around the face boundary.
• Early stopping + ReduceLROnPlateau — Best weights are restored automatically if validation loss degrades. The learning rate halves during Phase 2 if validation loss stalls for 2 epochs.

---

🛠️ Tech Stack

Component	Library
Neural network	TensorFlow 2.x / Keras
Pretrained backbone	MobileNetV2 (ImageNet weights)
Face detection	MediaPipe (BlazeFace)
Data pipeline	tf.data
UI	Gradio Blocks
Image processing	Pillow, NumPy, OpenCV
Evaluation	scikit-learn, Matplotlib, Seaborn

---

🗂️ Dataset

Face Mask Detection by Chandrika Deb

Split	With Mask	Without Mask
Train (80%)	552	549
Val (20%)	138	137
Total	690	686

The dataset is balanced — no class weighting needed.

---

🚀 How to Run

Google Colab (Recommended)

1. Upload face_mask_detector.ipynb to Google Colab
2. Enable GPU: Runtime → Change runtime type → T4 GPU
3. Runtime → Run all (Ctrl+F9)
4. In the last cell, a Gradio public URL will appear — open it and upload any photo

Local Machine

pip install tensorflow gradio mediapipe opencv-python-headless seaborn pillow scikit-learn

jupyter notebook face_mask_detector.ipynb

Note: Remove share=True from app.launch() when running locally — it's only needed in Colab to generate the public tunnel URL.

---

💡 What I Learned

• Transfer learning — reusing pretrained ImageNet weights and adapting them to a new task
• Two-phase fine-tuning — when to freeze layers, when to unfreeze them, and why the learning rate must drop before you do
• tf.data pipelines — building fast, parallelized data loading that doesn't bottleneck GPU training
• MediaPipe — using a neural face detector instead of Haar Cascades, and understanding the difference in robustness
• Model-integrated preprocessing — why embedding preprocess_input inside the model graph is safer than doing it externally
• Multi-model pipelines — chaining a detection model (MediaPipe) with a classification model (MobileNetV2)
• Evaluation — reading a confusion matrix and classification report, not just a single accuracy number

---

👤 Author

Mohamed Ouledali — Engineering Student