Sistema automático de predicción estética basado en Computación Evolutiva y Deep Learning

Abstract

Nowadays, with the rise of social media, we are used to making decisions based on the aesthetic value of images. In e-commerce, for example, we make purchasing decisions based on product images. In this context, an automatic system that allows us to select and sort images according to their aesthetic value can be of great value. This thesis addresses different problems in the field of computational aesthetics and proposes new solutions that are finally validated in a real case study. First, we study the datasets used in computational aesthetics and propose a new methodology for the creation of generalisable image sets that can be applied to Machine Learning problems. Then, a new approach using transfer learning with a new hybrid genetic algorithm for the prediction of aesthetic value in digital images is presented. Finally, both the proposed dataset creation methodology and the hybrid model that gave the best results in the experimental phase are applied to a real case study. The results suggest that using these tools can improve both user experience and e-commerce productivity.