Classification of art pictures using transformers

This thesis explores the classification performance of artworks based on three key attributes: "Genre," "Artist," and "Style," employing deep learning techniques. Convolutional neural network (CNN)-based methods have traditionally shown strong results in art classification tasks due to their effectiveness in feature extraction. However, transformer-based models, which utilize attention mechanisms, have recently gained prominence in the computer vision domain by demonstrating superior performance. In this study, classification experiments were conducted using several CNN-based models, including ResNet, MobileNet, EfficientNet, and ConvNeXt, alongside transformer-based models such as ViT, Swin, BEiT, and DeiT. These experiments utilized the publicly available WikiArt dataset. The highest classification accuracy in the "Artist" and "Genre" categories was achieved with the BEiT model at 84.90% and 79.52%, respectively, while the best result in the "Style" category was obtained with the Swin model at 72.59% accuracy. The findings indicate that transformer-based approaches consistently outperform their CNN counterparts. Additionally, when comparing these results with previous studies in the literature, it became evident that transformer-based models offer enhanced performance in the classification of artworks.