Large-scale photographic datasets like ImageNet and Places365 have significantly improved scene classification performance in natural images. However, scene classification in artistic imagery remains underexplored until recently. We propose a multi-step transfer learning technique that gradually adapts scene recognition algorithms from photographs to artistic scene representations. Our experiments demonstrate that integrating a stylized version of Places365, and fine-tuning with a weakly supervised artistic scene dataset, drastically increases scene recognition performance in artworks.We evaluate our method using two state-of-the-art scene recognition methods and analyze the impact of our adaptations with a series of ablation studies.