Discussion and Future

By doing this project, we have learnt how the CNN works and how to use CNN as a feature extractor for images. And how to design the loss function of the specific goal. Besides the designing part, we have also learnt how to implement a neural network with tensorflow and python.

We also tried to use our project to do video style transfer. But it is only theoretically approachable for us because of the running time. Our method transfers the input video frame by frame, but we believe there exists a cleverer approach that generate a style matrix as the continuous frames tend be similar.

In the future, we want to try to implement a general loss function that has good performance on any input and reference image for the CNN model. And recently, we find that Generative Adversarial Networks (GAN) has even more incredible result than those generated by CNN due the generative nature of this problem. So in the future, we want to try to use GAN to gain better performance on this topic. Some results generated by GAN:

And we want to have a speed-up method that can generate the output image more quickly so the result can be obtained in real-time. Then we can have a software with GUI, otherwise, generating a result with 1 hour is not possible for users.