You are on page 1of 8

Facial Recognition with

k-Nearest Neighbor and


Convolution Neural Network
Wyatt Steen
Project Overview

Goal:

- Train a machine learning model to distinguish and label five different faces.

- This project was intended to recognize faces of friends and family members, to deploy onto a Raspberry Pi

for a separate project.

- The model was trained using the faces of 5 different celebrities for discretion.

Research for a reason.


Datasets

- Celebrity-Face-Recognition-Dataset
- A dataset of 1100 Famous Celebrities with
700-800 thousand images each.
- Only 5 celebrities of different ages and races
were chosen.
- Anthony Mackie, Betty White, Beyonce
Knowles, Benidicio Del Toro, and Ben
Affleck
- Link to the dataset:
https://drive.google.com/drive/folders/0B5G8
pYUQMNZnLTBVaENWUWdzR0E?usp=sha
ring

Research for a reason.


Preprocessing
1. Detect faces in the image and use as the region of interest.

2. If more than one face is in the image, delete the observation.

3. If only one face detected, crop the region of interest.

4. Convert the cropped images to grayscale

5. Images sets for each of the five celebrities were then truncated to match the set with the least images.

Research for a reason.


Feature Extraction
1. HOG Features (Histogram of Oriented Gradient)
-Detects edges and orientation of the edges
2. Harris Features
-A corner Detector

3. SURF (Speeded up Robust Features)


-Wavelet (like Fourier Transform but more confusing) responses in X and Y directions

HOG Harris SURF

Research for a reason.


Models Investigated
- K-Nearest Neighbor

- Convolution Neural Network

Research for a reason.


Model Performance Comparison
- KNN Confusion Matrix and Derived Metrics

Accuracy Recall Specificity Precision


________ ______ ___________ _________
KNN Metrics 0.76744 0.9 0.65217 0.69231

Research for a reason.


References
[1] Prateekmehta59. (2018). Prateekmehta59/Celebrity-Face-Recognition-Dataset. Retrieved
November 25, 2020, from
https://github.com/prateekmehta59/Celebrity-Face-Recognition-Dataset

Thank You!

Research for a reason.

You might also like