Object Detection in Artworks Using Data Augmentation

Object detection is useful for authenticating artworks, and object detection methods based on deep learning have been actively studied in recent years. However, existing object detection methods cannot perform well on artworks, due to the lack of labeled artwork datasets. To overcome the problem, we propose a novel approach for object detection on artworks. The proposed approach consists of two steps. First, we apply neural style transfer to the existing images for data augmentation. Second, we train an object detection method on the augmented dataset. We evaluate the proposed approach on Brueghel and People-Art datasets. The experimental results show that our proposed approach successfully detects objects in artworks. Moreover, our approach achieved 40.6 mAP, which is 7.8 mAP higher than the conventional approach without data augmentation. It is expected that the proposed approach is of great benefit to art historians for analyzing a large number of artworks.