Host: The Institute of Image Electronics Engineers of Japan
Name : Reports of the276th Technical Conference of the Institute of Image Electronics Engineers of Japan
Number : 276
Location : [in Japanese]
Date : March 03, 2016 - March 04, 2016
With drastic increase in the collection of digitalized artworks, archiving the artworks database by human become extremely difficult. Hence, the aid of machine learning technique is important to automate such task. Recently, Convolutional Neural Network (CNN) has become a popular choice for feature extraction and classification. Many studies indicate that features learnt by CNN through ImageNet can be generalized to many other similar tasks. This paper demonstrates that such CNN features can even be generalized to fine-art paintings classifications. In this work, we focus on Style, Genre, and Artist classifications problems. We used the recently available large-scale Wikiart paintings dataset that is publicly available. We show that an end-to-end trained CNN is able to achieve descent performance for these tasks. More importantly, an ImageNet pre-trained CNN is able to achieve better recognition. In addition, fine-tuning allows the pretrained CNN to better adapt to the new tasks, resulting in significant improvement. Furthermore, we observe that Softmax Regression and Support Vector Machine performed on par in our experiments.