Approach to Minimize Bias on Aesthetic Image Datasets

Abstract

Over the last few years, numerous studies have been conducted that have sought to address automatic image classification. These approaches have used a variety of experimental sets of images from several photography sites. In this chapter, the authors look at some of the most widely used in the field of computational aesthetics as well as the capacity for generalization that each of them offers. Furthermore, a set of images built up by psychologists is described in order to predict perceptual complexity as assessed by a closed group of persons in a controlled experimental setup. Lastly, a new hybrid method is proposed for the construction of a set of images or a dataset for the assessment and classification of aesthetic criteria. This method brings together the advantages of datasets based on photography websites and those of a dataset where assessment is made under controlled experimental conditions.

Key Terms in this Chapter

Information Bias: Referred to as observational bias and misclassification.

Generalization: Capability to predict outcome values for previously unseen data.

Entropy: Lack of order or predictability; gradual decline into disorder.

Visual Perception: The ability to interpret the surrounding environment using vision.

Intrinsic Bias: Subconscious stereotypes that affect the way we make decisions.

Extrapolate: To use existing information to discover what is likely to happen or be.

Cross-Validation: A model validation technique for assessing how the results of a statistical analysis will generalize to an independent data set.