Face Detection Project using CNN

By studying machine learning concepts and architectures, I developed a simple project to test and enhance my knowledge in the field. This project involves creating a program that detects faces in real-time using a machine learning algorithm. The algorithm collects images from the user and trains them using a Convolutional Neural Network (CNN), enabling real-time classification through the webcam.

Table of Contents

• General Overview
• Architecture Overview
• Summary
• Dependencies
• Installation
• Usage
• License

General Overview

For building this model, I used PyTorch, which allowed me to manually manage the model and to learn in practise how to build from scratch. Visual representation of the CNN architecture.

Architecture Overview

1. Input: 120x120x3 (RGB Image)

2. Conv Layer 1:
   • Convolution: 16 filters, 3x3 kernel, stride=1, padding=1
   • Batch Normalization + ReLU
   • Max Pooling: 2x2, stride=2
   • Output: 60x60x16

3. Conv Layer 2:
   • Convolution: 32 filters, 3x3 kernel, stride=1, padding=1
   • Batch Normalization + ReLU
   • Max Pooling: 2x2, stride=2
   • Output: 30x30x32

4. Conv Layer 3:
   • Convolution: 64 filters, 3x3 kernel, stride=1, padding=1
   • Batch Normalization + ReLU
   • Max Pooling: 2x2, stride=2
   • Output: 15x15x64

5. Conv Layer 4:
   • Convolution: 128 filters, 3x3 kernel, stride=1, padding=1
   • Batch Normalization + ReLU
   • Max Pooling: 2x2, stride=2
   • Output: 7x7x128

6. Flatten: The 7x7x128 output is flattened into a 1D vector of size 6272.

7. Fully Connected Layer 1:
   • Input: 6272
   • Output: 512 neurons

8. Fully Connected Layer 2 (Output Layer):
   • Input: 512 neurons
   • Output: 5 neurons, 1 channel for the classifier (0, 1... It depends on how many subjects) and 4 channels for the Boundaries Boxes.

Summary

• The network progressively reduces the spatial size through max pooling while increasing the depth with more filters.
• After the convolutional layers, the features are flattened and passed through fully connected layers for classification.
• Finally, after training, the model is loaded and images taken in real time, by the webcam, are passed through the model resulting into a classifier and a Bbox.

Result of the model working properly identifying me and drawing the Boundary Box.

Dependencies

For running the Notebook correctly you will need to install the following Dependencies:

1. Install PyTorch:

https://pytorch.org/get-started/locally/

2. Install OpenCV:

 pip install opencv-python

3. Install LabelMe:

This part will be important because LabelMe does all the process of classifying your images and generating labels for them much more easily:

• LabelMe

Usage

1. To run the Notebook, I recomend using Jupyter Lab itself:

pip install jupyterlab

Or You can use any IDE of your own choice.

2. Follow the steps inside the Notebook.

License

This project is licensed under the Apache License 2.0.