Face Detector

Detect faces in images through the use of Transfer Learning.

• Use Deep Learning techniques CNNs, DNNs, and Transfer Learning.
• Graph training and validation accuracy to visualize the accuracy loss to prevent overfitting.
• Test facial verification using Euclidean and Cosine distance metrics on photo similarity matrices.

Technologies used: Densely Connected Networks, Convolutional Neural Networks Transfer Learning Libraries: MobileNet, InceptionNet, VGG16

Obtain

Import images:

• 3064 pictures of human faces from faces_data Kaggle DataSet.

• 1377 non-human faces and miscellaneous images.

Scrub and Clean the DataSet

• Rescale images to the same dimensions with keras.preprocessing.image library.
• Split the images to train our Supervised Machine Learning classification models.

Training Models

Densely Connected Network (DCN)

• Adding two hidden layers, testing different numbers of nodes and activation functions.
• Optimize using SGD, and binary_crossentropy for thirty epochs which yielded a 99.86% accuracy. Tools used: Tensorflow, SGD, binary_crossentropy

Convolutional Neural Network

• Achieve an F1 score of 100%.
• Use convolutional neural network to input sequential layers.
• Mix node equations in different ways for each layer.

Drop-out Regularization - Addressing overfitting

• Remove 50% of the nodes from the subsequent layer allowing for a more robust interpretation of neural network.
• Train accuracy went down and our F1 scores went down.
• Achieve more robust, neural network.

Transfer Learning

Using Transfer Learning, replace the last layer of the neural network to classify and differentiate your target in the images.

• Transfer Learning models used Deep Learning Neural Networks that have been trained on millions of images. Explanation: Models have been tuned on millions of pictures, the weights of each of these nodes have captured robust nuances of the intended target within the photos.

MobileNet

InceptionNet

VGG16 here

• Use a sigmoid function to determine if the image showed a face or not for the last layers.

Evaluating Models:

• Achieve an F1 Score of 99.84%
• Compare training and validation loss functions against training and validation accuracy to ensure good fit.
• Fit at 60 epochs, went down to 10 epochs.

Implement

• Test individual images using the model.

Summary

The best results were achieved through Transfer Learning library VGG16.

Next Steps

• Incorporate video.
• Facial Verification.
• Implement on Arduino project.