Tools
Tools
Kori, Yogaraj Chandrashekar (2024) Fictional Face Generation using DCGAN and WGAN with Gradient Penalty. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (1MB) | Preview |
Preview |
PDF (Configuration Manual)
Download (1MB) | Preview |
Abstract
With the increase in the capability of Artificial Intelligence and Machine Learning Algorithms, scams and deep-fake profiles on social media websites are becoming predominant. As such it is important to counter such malicious activities by bettering existing human face detection algorithms and testing them with deep-fake images of our own. In this project, we will be providing a human face generation algorithm with the ability to create high end facial images of non-existent personalities using Deep Convolutional Generative Adversarial Networks (DC-GAN) and Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN + GP). A collection of thirty thousand images of cropped and augmented human faces was used to train the model where two parts of the algorithm generator and discriminator compete with each other in creation and challenging of new images. To analyze the model performance, we will be monitoring Inception Score (IS) of the generated images along with plotting the generator and the discriminator loses to understand the imbalance. Along with Inception Score we have also utilized Structural Similarity Index (SSIM) which compares the contextual and structural similarity of generated images and its training image counterparts as well as Fréchet Inception Distance (FID). With this implementation we can generate non-existent human faces and utilize it to further improve facial recognition software by feeding these deep-fake images for negative validations and edge case scenario validation.
Actions (login required)
 |
View Item |