How Fake is my Image? Evaluation of Generative Adversarial Networks

September 23, 2021

Generative machine learning is a very recent field and has been proven very successful in many multimedia applications. One that has had much impact is the creation of fake images or video clips which can challenge the human perceived notion of reality. Generative Adversarial Networks (GANS) have risen in popularity among the generative models that can learn high-dimensional distributions of data (i.e. the manifold of the entire set of natural images). And have been applied successfully to several tasks such as image compression. Naturally, for these solutions is important to assess the perceptual quality of generative images which will be shown to the user. However, it was recently shown that the typical objective Image Quality Assessment (IQA) metrics have proven to be inefficient at evaluating the perceptual quality of GAN produced images. Also, other quality factors play a role besides fidelity, such as the naturalness/fakeness of the image. Here is presented a subjective and objective study of image quality of GAN created images.

Authors in the community:

• 

  M

  Marta Filipa De Pinto Marques

  ist180882

Supervisors of this institution:

• 

  C

  Catarina Isabel Carvalheiro Brites Ascenso

  ist146502

• 

  J

  João Miguel Duarte Ascenso

  ist31104

electrical-engineering-electronic-engineering-information-engineering - Electrical engineering, electronic engineering, information engineering

eng - English

Instituto Superior Técnico