Real-Time Face Mask Detection with Deep Learning

Objective: Detect whether a person is wearing a face mask in real time via webcam — by combining face detection and image classification using a neural network.

This project implements a complete face mask detection system:

1. Model Training: Image classification model (using MobileNetV2) trained on a dataset of faces with and without masks.
2. Real-Time Detection via webcam: Face localization + prediction of “Mask” / “No Mask” with probability.

Ideal for health safety applications, access control systems, or educational AI projects.

---

Target Audience

Audience	What They Will Find
Students in AI / Computer Vision	A complete tutorial, from training to real-time detection, with clear code and explanations.
Teachers / Trainers	An ideal pedagogical resource for teaching transfer learning, data augmentation, and real-time inference.
Developers / Data Scientists	A clean implementation using TensorFlow/Keras, OpenCV, and imutils — easy to adapt, deploy, or improve.
Curious Non-Technical Users	An impressive and practical demo: see how AI can “see” and “understand” what you’re wearing on your face!

---

Key Features

1. Model Training (train_mask_detector.py)

• Dataset: Labeled face images (“with_mask” / “without_mask”)
• Base Model: MobileNetV2 (pre-trained on ImageNet) → lightweight and fast
• Fine-tuning: Custom classification head added:
  • AveragePooling2D
  • Flatten
  • Dense(128, relu)
  • Dropout(0.5)
  • Dense(2, softmax)
• Data Augmentation: Rotations, zooms, shifts, flips → improves generalization
• Optimization: Adam optimizer, lr=1e-4, learning rate decay
• Metrics: Accuracy, Loss (training + validation)
• Output: Saved model → mask_detector.model

2. Real-Time Detection (detect_mask_video.py)

• Face Detection: Caffe SSD model (res10_300x300_ssd_iter_140000.caffemodel)
• Mask Prediction: Load trained model (mask_detector.model)
• Real-Time Pipeline:
  1. Video capture (webcam)
  2. Detect all faces in frame
  3. For each face:
     • Extract and resize to 224x224
     • Preprocess (preprocess_input)
     • Predict “Mask” / “No Mask” + confidence probability
     • Display label and colored bounding box (green = mask, red = no mask)
• Control: Press ‘q’ to quit

---

Typical Results

• Validation Accuracy: > 95% (depending on dataset quality)
• Detection Speed: 15–30 FPS on a modern computer (CPU)
• Training Curves generated (plot.png) → monitor convergence

---

Technologies & Libraries

# Training
from tensorflow.keras.applications import MobileNetV2
from tensorflow.keras.layers import AveragePooling2D, Dropout, Flatten, Dense
from tensorflow.keras.optimizers import Adam
from sklearn.preprocessing import LabelBinarizer
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report

# Detection
from tensorflow.keras.models import load_model
import cv2
from imutils.video import VideoStream
import numpy as np