Near Human-Level Style Transfer
Rahul Pereira, Beryl Coutinho, Jenslee Dsouza, Cyrus Ferreira, Vaishali S Jadhav · Auerbach Publications eBooks · 2023
The model can also work on photographs but the best results are obtained when it is provided with artworks. Specifically, three types of loss are considered, i.e. style loss, content loss, and variation loss. Using this approach, i.e. starting from zero, is like reinventing the wheel; instead, our research proposes using the concept of transfer learning wherein a pre-trained model is reused based on our use case and is fine-tuned as per our requirements. Since our research only requires feature extraction, we have excluded the remaining part of the VGG19 network. Style loss is similar to content loss, but the difference lies in that it is used to ensure the style of the style image is preserved in the final generated output image. Since L-BFGS is a very memory-expensive algorithm, our research proposes using gradient descent optimization technique to maximize the model performance.