Eigenfaces

Group: ASA Anastasiia Livochka, Sofia Kholod, Andriy Dmytruk
Goal: Use the Eigenfaces approach to recognize our groupmates

Usage

Server

To run the local server with face recognition, you need to start:

python main.py server [FOLDER] [DATASET_FOLDER]

This will start the server at localhost:4000. The parameter folder is not required and needs to have the path to a dataset of normalized people photos.
This will train the algorithm on dataset and will be ready to recognize.

Website supports taking photos online, uploading photos stored locally and switching camera.
Website can also be used to collect the dataset. The second parameter [DATASET_FOLDER] is the folder where all dataset photos will be saved. The endpoint localhost:4000/eigenfaces/save helps you easily take the photos and automatically save them to dataset.

Dataset & Normalization

The original dataset can be found in at Google Drive.
The normalized dataset is stored in normalized_apps folder. So, if you just want to test the recognition, you do not need to manually normalize the dataset.

Before running recognition, you need to normalize the photos in dataset.
To do this run:

python normalize_images.py [SOURCE_FOLDER] [RESULT_FOLDER]

The default parameters are ./apps_faces and ./normalized_apps

Testing

Testing is performed automatically after normalization to inform about how good the normalization performed and whether the images given were good enough.

It removes some of the photos from dataset, trains the algorithm without them and then calculates accuracy on these photos. This is repeated a few times and the average is shown to user.

Manual face recognition

Other examples of using the algorithm are given in main.py.
For example, you can generate images of all the eigenfaces with save_eigenfaces() or see you own images in the reduced basis with test_detection_procedure().

Note that all the functions work with already normalized photos.

Algorithm

Normalization

1. Face detection was performed using haar cascade and cv library. After detecting, faces where rescaled to 200x200 pixels.
2. Then gamma correction and histogram equalisation normalized the lighting on images.
3. A gabor filter was used to further make details more vivid.

Training

Training was performed using PCA.
For all the images in dataset, their weight (vector in the reduced vector space of eigenvectors of the covariance matrix) is calucated.

Recognition

To recognize the person on a picture, we compared the weight of face on image to the average weight of images of every person in dataset.

The comparison was performed based on the distance between the weight vectors. The person with the lowest distance was chosen.

To produce the percentages for each person, we decided to approximate the distance from person's weight vector to the average vector using normal distribution. Then we used the cdf to get the probability. If the distance is much larger than its expected value, the probability is low. If the distance is much lower, than its almost definitely that person.

Results

The algorithm produced accuracy of 81% on the dataset in the end and was able to detect the person in real-life conditions using the website.