Transferring artistic styles onto everyday photographs has become anextremely popular task in both academia and industry. Recently, offlinetraining has replaced on-line iterative optimization, enabling nearly real-timestylization. When those stylization networks are applied directly tohigh-resolution images, however, the style of localized regions often appearsless similar to the desired artistic style. This is because the transferprocess fails to capture small, intricate textures and maintain correct texturescales of the artworks. Here we propose a multimodal convolutional neuralnetwork that takes into consideration faithful representations of both colorand luminance channels, and performs stylization hierarchically with multiplelosses of increasing scales. Compared to state-of-the-art networks, our networkcan also perform style transfer in nearly real-time by conducting much moresophisticated training offline. By properly handling style and texture cues atmultiple scales using several modalities, we can transfer not just large-scale,obvious style cues but also subtle, exquisite ones. That is, our scheme cangenerate results that are visually pleasing and more similar to multipledesired artistic styles with color and texture cues at multiple scales.
展开▼
