A Small Scale Multi-Column Network for Aesthetic Classification Based on Multiple Attributes

EVA: An Explainable Visual Aesthetics Dataset

“…In [12], the authors have verified how high-level attributes, like style and semantic, affect aesthetic scores. In [23], authors use a multi-column neural network to first train different visual factors and semantic features, then combine them together with a column for unknown attributes, to imitate the general aesthetic values. As an intrinsically subjective task, the collection of reliable data labels is a significant challenge, which greatly impacts the development of effective models.…”

“…Exploiting these existing datasets, researchers have aimed at explaining visual aesthetics. In [11,23], attribute labels are computed from images instead of asking the subjects directly. As a consequence, the relative importance among attributes depend on the feature extraction models.…”

“…For each attribute, subjects have to vote on a four-level Likert scale with the following options: very bad, bad, good, and very good. We choose these four attributes, as they have been previously studied [1,9,10,23] and they are relatively easy to understand by naive subjects. The attributes have been defined as follows in the user training phase prior to the test: Light and color relates to visual perception, including brightness, contrast, and color saturation.…”

“…Previous research has studied the relationship between aesthetic attributes and aesthetic scores. A multi-column network is proposed in [23] to learn this relationship automatically, but the features rely on manual design, which may introduce a bias. The approaches in [2,24,26] aim at finding the relative importance of aesthetic attributes based on text comments.…”