We consider image transformation problems, where an input image istransformed into an output image. Recent methods for such problems typicallytrain feed-forward convolutional neural networks using a \emph{per-pixel} lossbetween the output and ground-truth images. Parallel work has shown thathigh-quality images can be generated by defining and optimizing\emph{perceptual} loss functions based on high-level features extracted frompretrained networks. We combine the benefits of both approaches, and proposethe use of perceptual loss functions for training feed-forward networks forimage transformation tasks. We show results on image style transfer, where afeed-forward network is trained to solve the optimization problem proposed byGatys et al in real-time. Compared to the optimization-based method, ournetwork gives similar qualitative results but is three orders of magnitudefaster. We also experiment with single-image super-resolution, where replacinga per-pixel loss with a perceptual loss gives visually pleasing results.
展开▼