With advances in artificial intelligence (AI) and the availability of comprehensive digital copies of visual art, analyzing visual arts automatically becomes feasible. Extracting discriminative representations of visual art plays a key role in visual arts analytics, which is different from normal image analytics and has different sets of challenges. In normal images, the semantic concepts are usually regarded as powerful representations. However, in visual arts images, the information is conveyed by the subject matter and the style. Style is not necessarily correlated with subject matter, which corresponds to the existence of certain objects or contents in the images. Therefore, a unified framework is proposed in this thesis to learn joint representations that can simultaneously...[ Read more ]

With advances in artificial intelligence (AI) and the availability of comprehensive digital copies of visual art, analyzing visual arts automatically becomes feasible. Extracting discriminative representations of visual art plays a key role in visual arts analytics, which is different from normal image analytics and has different sets of challenges. In normal images, the semantic concepts are usually regarded as powerful representations. However, in visual arts images, the information is conveyed by the subject matter and the style. Style is not necessarily correlated with subject matter, which corresponds to the existence of certain objects or contents in the images. Therefore, a unified framework is proposed in this thesis to learn joint representations that can simultaneously capture the content and style of visual art. In the real world, many applications (e.g., visual arts recognition, retrieval, recommendation, etc.) can be built based on the proposed learned representations of visual art. However, in most scenarios, the performance of the proposed framework will suffer from different issues (e.g., distortions, blur) existing in the images, especially for mobile-based applications. To overcome this problem, an image rectification pipeline with three modules is designed to rectify the images. The effectiveness of the proposed methods has been proved through a series of experiments. Additionally, a real-world system is implemented¹ and two visual arts datasets² are released to facilitate research in this area.

¹http://deepart2.ece.ust.hk

²http://deepart2.ece.ust.hk/ART500K/art500k.html