• artist classification
• cnn
• convolutional neural network
• machine learning
• paintings classification
• similarity search
• style classification

The purpose of this thesis is the development of a Web Application to categorize paintings and search for similarity by style. For this purpose, a Convolutional Neural Network (CNN) has been trained on two datasets, one of 13 style classes and one of 91 artist classes. Determining the style and artist of a painting can be difficult even for an expert, and sometimes it is also possible for two experts to express different opinions. Also, if it is possible to determine a unique artist for a painting, it is much more difficult to understand if the style is one, or if it has not been influenced by other styles. The challenge is to be able to extract the true style of the painting and to be able to identify its artist regardless of the period in which she/he made it. Another challenge is that the datasets available for this type of task are not large enough to allow the training of networks from scratch. In this thesis work, we used differents CNNs, with a training composed of two parts and the dataset Paintings91 was used. A part is based on the artist's classification, the other on style classification. For both we used a process of transfer learning to reduce the training time and to use a dataset not so big to allow the training from scratch. In addition, a Web Application was developed. Users can use images as query and obtain its style, its artist classification and similar images.The best accuracy is using residual networks.