README.md

Facial Recognition Using Deep Neural Network

This repository includes jupyter notebooks and test pictures of our project.

We are trying to predict the names of people in a picture provided.

To do this, we considered a well-known pre-trained Convolutional Deep Neural Networks, called VGG-FACE.

Then we created our own datasets and fine tuned the model, trained the customized model again.

Our Team Members are:

• Sui Huang: sh4507@nyu.edu
• Bowen Li: bl2305@nyu.edu

Facial Recognition Using Pre-Trained VGG-Face + OpenCV

Jupyter notebook of our project can be found here Pre_Trained_Vgg_Face.ipynb

In VGG-Face, the dataset has already be trained and weights can be downloaded from here.

We are building our project according to this page. We are using the OpenCV cascade to cut out faces referencing this page.

Result:

Train VGG-Face Architecture Using Our Own Dataset

Jupyter notebook of our project can be found here Fine_Tuning_Vgg_Face.ipynb.

We created our own dataset with 1496 photos of 10 different celebrities. The size of our databset is over 100 files so you can download it from here (link deleted).

UPDATE

Please do not ask for access to this dataset. This is a private dataset created by our own with huge efforts. Niether do I have the right to share nor will I share. Thanks.

We were trying to fine tune layer to VGG-Face achitecture refrencing this page and this page.

We also imported some twists to our dataset, and compared the results to the origin one.

Result:

Model No	Training Set	Testing Set	Accuracy	Validate Accuracy
1	Original	Original	>98%	>95%
2	Original	Twisted	>98%	>25%
3	Twisted	Twisted	>44%	>42%

Library Version

python 3.6.1

Anaconda 4.3.30

Tensorflow 1.4.0

Keras 2.0.9

Pillow 4.3.0

OpenCV-python 3.3.0.10

System Configuration

OS: Windows10

CPU: Intel(R) Core(TM) i7-6700

GPU: NVIDIA GeForce GTX 1070