Generative Adversarial Networks for Improving Face Classification

Abstract

Facial recognition can be applied in a wide variety of cases, including entertainment purposes and biometric security. In this thesis we take a look at improving the results of an existing facial recognition approach by utilizing generative adversarial networks to improve the existing dataset. The training data was taken from the LFW dataset[4] and was preprocessed using OpenCV[2] for face detection. The faces in the dataset was cropped and resized so every image is the same size and can easily be passed to a convolutional neural network. To the best of our knowledge no generative adversarial network approach has been applied to facial recognition by generating training data for classification with convolutional neural networks. The proposed approach to improving face classification accuracy is not improving the classification algorithm itself but rather improving the dataset by generating more data. In this thesis we attempt to use generative adversarial networks to generate new data. We achieve an impressive accuracy of 99.42% with 3 classes, which is an improvement of 1.74% compared to not generating any new data.