Explorando Padrões Artísticos em Pinturas Impressionistas: Uma Abordagem de Aprendizado Não Supervisionado

Abstract

At the confluence of art and science, the volume of research exploring the application of artificial intelligence techniques in the analysis of works of art has grown rapidly. This study sought to contribute to this expanding field, focusing specifically on the computational analysis of impressionist paintings. Faced with the vastness and complexity of artistic works, the task of extracting meaningful knowledge represents a significant computational challenge. The implementation present in this study proceeded with the reduction of dimensionality of the features extracted from 304 impressionist paintings with PCA and t-SNE, and clustering methods such as k-means and hierarchical clustering. Using the elbow method, it was concluded that five was the ideal number of clusters, both for PCA and for t-SNE. This f inding was reinforced by the silhouette coefficient, which evidenced superior cohesion for the k-means clustering with k = 5. The results also showed a unique evaluation with respect to genre, with the PCA method and k-means with k = 5 showing a higher genre entropy, whether with 1,868 for unnormalized data or 2,341 for normalized data. Similarly, this clustering method led the entropy of artists, with a value of 4,036 for unnormalized data and 4,186 for normalized data. However, in the case of t-SNE, although k-means with k = 5 also demonstrated superiority in the entropy values of years and artists. Still, the general entropy metric of t-SNE corroborated the efficacy of k-means clustering with k = 5.