Abstract
Abstract
Automatic painting classification by author, artistic genre and/or other attributes has generated considerable research interest in recent years. Being one of the visual features that mark the difference between artists and artistic genres, colour plays a fundamental role in this process. Colour is the result of the interaction among the intrinsic properties of the material, the illumination conditions and the response of the imaging device. Consequently, the same painting/artwork will look significantly different when imaged under varied conditions, which can be a potential source of bias for automated recognition procedures. One can compensate for such variations either via colour calibration or colour pre-processing. In this work we investigate the latter, and, in particular, evaluate the effectiveness of colour constancy and colour augmentation when coupled with hand-crafted and deep learning features for painting classification by artistic genre. In our experiments neither approach showed a clear advantage compared with no pre-processing at all. Colour constancy brought some improvement in certain cases, whereas colour augmentation virtually provided no benefit despite its adding a significant computational overload to the procedure.