Disentanglement of Human Facial Features and the Inspection of Feature Extractor’s Quality Based on TL-GAN

Abstract

Recent research on the Generated Adversarial Network models has made great progress in the task of generating images on human demand. Among them, Transparent Latent-space GAN tries to solve the problem by analyzing the latent space and finding the feature axes inside it. This model views the problem from a novel perspective, and it is easy to be carried out. However, more thorough and detailed experiments have shown that the original method on axes disentanglement of the model has certain drawbacks, which may lead to ethical problems, meaning that the model may pick up some pre-existing prejudice of racial or gender discrimination. This paper puts forward a new training mode aiming at reducing the bias existing in the generated images and simultaneously managing the disentanglement. Also, a more thorough examination of the model was conducted and it is found that the architecture of the model may be generalized to some universal functions. Among them, the quality evaluation of the feature extractor is the most practical and useful one. This finding, in turn, may be helpful for distinguishing a more accurate, effective, and robust feature extractor, thus improving the performance of the TL-GAN model in the first place.