Clasificación de estilos pictóricos utilizando redes neuronales convolucionales

Abstract

The main objective of this project was to classify paintings of diverse styles by the artistic style using convolutional neural networks. These types of networks are deep learning models that have been widely used in image classification tasks. In this project, three different models were trained and tested with a dataset containing 16 different artistic styles: a simple network, the VGG-16 network and the ResNet-50 model. Before starting the multiclass classification experiment, a binary classification problem that aimed to determine how well those networks differentiated two artistic styles (Baroque and Impressionism) was evaluated with those convolutional models.

Besides, the three types of networks mentioned have also been used to solve another image classification task that consisted of distinguishing photographs from paintings. The three models were retrained with a dataset containing paintings of various styles and photographs ranged from amateur level to professional level.

Another task that has been solved in this project was to study and analyze the neural style transfer algorithm. Along with that, the original convolutional model for which the style transfer was proposed was replaced with one of the networks that had been retrained in this project to investigate how different the results would be with that modified model.

Finally, a graphical user interface was built to test the three different experiments in a more user-friendly way. It is intended for facilitating the application of the models without modifying the code.