A system for modeling social traits in realistic faces with artificial intelligence
- Autores: Félix José Fuentes Hurtado
- Directores de la Tesis: Valeriana Naranjo Ornedo (dir. tes.), José Antonio Diego Mas (dir. tes.)
- Lectura: En la Universitat Politècnica de València ( España ) en 2018
- Idioma: español
- Tribunal Calificador de la Tesis: José Javier López Monfort (presid.), Martin Risdal (secret.), Antonio González Pardo (voc.)
- Programa de doctorado: Programa de Doctorado en Telecomunicación por la Universitat Politècnica de València
- Materias:
- Enlaces
- Tesis en acceso abierto en: RiuNet
- Resumen
Humans have specially developed their perceptual capacity to process faces and to extract information from facial features. Using our behavioral capacity to perceive faces, we make attributions such as personality, intelligence or trustworthiness based on facial appearance that often have a strong impact on social behavior in different domains. Therefore, faces play a central role in our relationships with other people and in our everyday decisions.
With the popularization of the Internet, people participate in many kinds of virtual interactions, from social experiences, such as games, dating or communities, to professional activities, such as e-commerce, e-learning, e-therapy or e-health. These virtual interactions manifest the need for faces that represent the actual people interacting in the digital world: thus the concept of avatar emerged. Avatars are used to represent users in different scenarios and scopes, from personal life to professional situations. In all these cases, the appearance of the avatar may have an effect not only on other person's opinion and perception but on self-perception, influencing the subject's own attitude and behavior. In fact, avatars are often employed to elicit impressions or emotions through non-verbal expressions, and are able to improve online interactions or even useful for education purposes or therapy. Then, being able to generate realistic looking avatars which elicit a certain set of desired social impressions poses a very interesting and novel tool, useful in a wide range of fields.
This thesis proposes a novel method for generating realistic looking faces with an associated social profile comprising 15 different impressions. For this purpose, several partial objectives were accomplished.
First, facial features were extracted from a database of real faces and grouped by appearance in an automatic and objective manner employing dimensionality reduction and clustering techniques. This yielded a taxonomy which allows to systematically and objectively codify faces according to the previously obtained clusters. Furthermore, the use of the proposed method is not restricted to facial features, and it should be possible to extend its use to automatically group any other kind of images by appearance.
Second, the existing relationships among the different facial features and the social impressions were found. This helps to know how much a certain facial feature influences the perception of a given social impression, allowing to focus on the most important feature or features when designing faces with a sought social perception.
Third, an image editing method was implemented to generate a completely new, realistic face from just a face definition using the aforementioned facial feature taxonomy.
Finally, a system to generate realistic faces with an associated social trait profile was developed, which fulfills the main objective of the present thesis.
The main novelty of this work resides in the ability to work with several trait dimensions at a time on realistic faces. Thus, in contrast with the previous works that use noisy images, or cartoon-like or synthetic faces, the system developed in this thesis allows to generate realistic looking faces choosing the desired levels of fifteen impressions, namely Afraid, Angry, Attractive, Babyface, Disgusted, Dominant, Feminine, Happy, Masculine, Prototypical, Sad, Surprised, Threatening, Trustworthy and Unusual.
The promising results obtained in this thesis will allow to further investigate how to model social perception in faces using a completely new approach.
