Classifying images by light aesthetics - DUMAS - Dépôt Universitaire de Mémoires Après Soutenance Accéder directement au contenu
Mémoires Année : 2013

Classifying images by light aesthetics

Résumé

With the growing availability of storage space for data, such as images, comes the need for research tools that reach beyond usual criteria. Such tools include, but are not limited to, the automatic prediction of aesthetic value for indexing purpose. For this reason, the field of supervised learning for aesthetic prediction has been widely active for more than a decade. We propose a new method aimed at classifying photos based on the aesthetic of their lightning. It comes as an introduction to providing more accurate specialized supervised learning solutions for inferring photo aesthetic. In the field of Aesthetic prediction classifier are broadly used, combining them with expert knowledge of the medium allow predictions with a score up to 70% matching the aesthetic judgement of humans. However, Supervised Learning solutions have a hard time dealing with the notion of different style in the datasets. While some aspects of photography, such a spatial composition, are considered as general and always applicable, the notion of style de fines specific aesthetics choice that change drastically the way the lightning (for example) is done. For this reason, the classifiers performances are limited. We choose to focus on the subject of light aesthetic in photography because we think it is relevant to be able to assess if a photo belongs to High-Key, Medium-Key or Low-key style prior to trying to infer it's aesthetic value. We show that it is possible to achieve such classification with over 80% success rate while only studying pixel populations of input images.
Fichier principal
Vignette du fichier
JulienSicre.pdf (5.62 Mo) Télécharger le fichier
Loading...

Dates et versions

dumas-00854874 , version 1 (28-08-2013)

Identifiants

  • HAL Id : dumas-00854874 , version 1

Citer

Julien Sicre. Classifying images by light aesthetics. Machine Learning [cs.LG]. 2013. ⟨dumas-00854874⟩
272 Consultations
294 Téléchargements

Partager

Gmail Mastodon Facebook X LinkedIn More