基于深度学习的图像风格迁移方法对比研究

A Comparative Study of Image Style Transfer Methods Based on Deep Learning

采用视觉质量、风格化时间、结构相似性等指标,对7种基于深度学习的图像风格迁移方法的性能进行了定性和定量比较,结果表明:基于图像迭代的神经风格迁移方法NST、松弛最优传输和自相似性STROTSS生成风格化图片视觉质量高,但速度慢且不稳定.对速度要求高时可使用基于模型迭代的方法,其中单网络单风格的快速风格迁移RTST和单网络多风格的条件实例归一化CIN生成风格化图片自然平滑,但不够灵活;对灵活性和功能性要求高时可使用单网络任意风格的自适应实例归一化AdaIN、特征变换WCT及注意力感知多笔画AAMS.

Using the indicators of visual quality,style transfer time and structural similarity,the performances of seven image style transfer methods based on deep learning were compared qualitatively and quantitatively.The results showed that the stylized images generated by the image-based iteration methods,including NST,STROTSS,had high but unstable visual quality with slow speed.When the speed requirement was high,the model-based iteration methods could be employed.RTST with single network and single style,as well as CIN with single network and multiple styles,could generate stylized pictures naturally and smoothly,but not flexible enough.When higher flexibility and functionality were required,single network arbitrary style method of AdaIN,WCT,and AAMS could be used.