Abstract

Following the rise of neuroaesthetics as a research domain, computational aesthetics has also known a regain in popularity over the past decade with many works using novel computer vision and machine learning techniques to evaluate the aesthetic value of visual information. This thesis presents a new approach where low-level features inspired from the human visual system are extracted from images to train a machine learning-based system to classify visual information depending on its aesthetics, regardless of the type of visual media. Extensive tests are developed to highlight strengths and weaknesses of such low-level features while establishing good practices in the domain of study of computational aesthetics. The aesthetic classification system is not only tested on the most widely used dataset of photographs, called AVA, on which it is trained initially, but also on other photographic datasets to evaluate the robustness of the learnt aesthetic preferences over other rating communities. The system is then assessed in terms of aesthetic classification on other types of visual media to investigate whether the learnt aesthetic preferences represent photography rules or more general aesthetic rules. The skill transfer from aesthetic classification of photos to videos demonstrates a satisfying correct classification rate of videos without any prior training on the test set created by Tzelepis et al. Moreover, the initial photograph classifier can also be used on feature films to investigate the classifier’s learnt visual preferences, due to films providing a large number of frames easily labellable. The study on aesthetic classification of videos concludes with a case study on the work by an online content creator. The classifier recognised a significantly greater percentage of aesthetically high frames in videos filmed in studios than on-the-go. The results obtained across datasets containing videos of diverse natures manifest the extent of the system’s aesthetic knowledge. To conclude, the evolution of low-level visual features is studied in popular culture such as in paintings and brand logos. The work attempts to link aesthetic preferences during contemplation tasks such as aesthetic rating of photographs with preferred low-level visual features in art creation. It questions whether favoured visual features usage varies over the life of a painter, implicitly showing a relationship with artistic expertise. Findings display significant changes in use of universally preferred features over influential vi abstract painters’ careers such an increase in cardinal lines and the colour blue; changes that were not observed in landscape painters. Regarding brand logos, only a few features evolved in a significant manner, most of them being colour-related features. Despite the incredible amount of data available online, phenomena developing over an entire life are still complicated to study. These computational experiments show that simple approaches focusing on the fundamentals instead of high-level measures allow to analyse artists’ visual preferences, as well as extract a community’s visual preferences from photos or videos while limiting impact from cultural and personal experiences.

Keywords

Visual preferences, Deep Learning, Aesthetic classification, Neuroaesthetics, Feature extraction

Document Type

Thesis

Publication Date

2018

Download

Share

COinS