Hierarchical classification of fine-art paintings using deep neural networks
- PDF / 957,432 Bytes
- 8 Pages / 595.276 x 790.866 pts Page_size
- 95 Downloads / 247 Views
ORIGINAL ARTICLE
Hierarchical classification of fine-art paintings using deep neural networks Mohammad Reza Mohammadi1
· Fatemeh Rustaee1
Received: 29 February 2020 / Accepted: 19 September 2020 © Springer Nature Switzerland AG 2020
Abstract Automatic classification, indexing, and retrieval of fine-art paintings have received much attention in recent years. Compared to the classification of natural objects, the classification of artworks is very challenging. To improve the performance of the deep neural networks for the style classification of artworks, we propose a hierarchical classification approach. In this approach, we group similar styles into several super-styles called parents. Then, we design one parent classifier and several child classifiers to classify the super-style as well as the style. Using this approach, we break down the complex problem of style classification into several simpler problems. The experimental results on the WikiArt dataset shown that the hierarchical classification approach improves the average F1 score of the DenseNet121 network from 55.7 to 59.1%. Keywords Deep neural networks · Convolutional layers · Fine-art paintings · Hierarchical classification · Style classification
1 Introduction The rapid advance of technology in the digitalization of artworks has led to a significant increase in the number of online art collections available. To manage and index these vast collections, we require to classify the paintings. Art experts can identify the style of paintings using their experience and knowledge of specific features in the artwork. However, doing this manually is very time consuming and requires historical and artistic expertise. Therefore, automatic recognition of artistic style using computer vision and machine learning techniques has attracted the attention of researchers. Deep learning, especially deep neural networks with convolutional layers, is the state-of-the-art in many fields of computer vision, including visual object recognition and semantic segmentation. In recent years, deep learning has been used successfully to classify artworks [2,13,26]. To train a deep neural network with millions of parameters, utilizing a large labeled dataset is essential. However, the available artworks datasets are not large enough. Therefore, in most studies on artworks classification, fine-tunning of the pre-
B 1
Mohammad Reza Mohammadi [email protected] School of Computer Engineering, Iran University of Science and Technology, Tehran, Iran
trained networks on large datasets such as ImageNet [9] is used. Compared to visual object recognition, there are some challenges in the classification of the art style. Among the artwork imagery, the analysis of portraits and landscapes are relatively simple. Still, the analysis of abstract paintings that use non-representational and non-allegorical forms requires artistic expertise and visualization. One of the most important challenges in the analysis of artworks is the ability of the machine to understand visualization and abstrac
Data Loading...
