Deep learning for object detection in fine-art paintings

Abstract

We propose deep learning and neural networks to automatically detect objects in digital pictures of fine-art paintings. This automatic annotation of digitized artwork provides innovation for content analysis, and therefore enhances the process of documenting and managing cultural heritage. Deep neural networks have outperformed all previous machine learning techniques in computer vision and achieve the highest accuracy in object detection. However, a very big amount of labeled training samples are required for such good performance. Typically, this big data is collected from everyday natural images, which is possible because millions are generated each day. Unfortunately there are not such big datasets of digitized fine-art paintings. In this contribution we present a set of strategies to overcome the lack of labeled training data, and hence make use of the promising deep learning in this application.