Neural style transfer: Can anyone become Van Gogh?

---

Neural Style Transfer(NST) is the technique of recomposing one image in the style of another. Two inputs, a content image and a style image are analyzed by a convolutional neural network which is then used to create an output image whose “content” mirrors the content image and whose style resembles that of the style image. It was first demonstrated in the paper “A neural algorithm of artistic style” published by Gatys, Ecker, and Bethge at the University of Tübingen in August 2015, and has since continued to be of great interest to artists and scientists alike. High-level neural networks capture the style of the picture, such as color and arrangement. The underlying neural network captures the content of the image, which is the details of the image.This is just the first step in a neural style transfer. After grabbing the style, iterate over the style to the image that needs to be changed. Style iteration usually has two ways, one based on the image, directly update the iterative image pixels, and finally achieve the style of migration. Many algorithms calculate the maximum mean difference in the process and measure the difference between the style image and the content image. Let the two images “align” to reduce the loss and error caused by image iterations.The other is based on model iteration. When a large number of images need to be iterated in a certain style, the feedforward network can be trained, and the network model can be optimized by iteratively updating the model using gradient descent.
Combine the content of the original photograph depicting the Neckarfront in Tubingen, Germany,with the style of a range of well-known artworks (such as The Shipwreck of the Minotaur, The Starry Night, Der Schrei) to create new images. The separation of image content and style is achieved through an artificial nervous system, allowing the content of an image to be recreated in any other image style.This is the first demonstration of image features separating content from style in whole natural images.
NST can save a lot of time in the art commercialization process, but the neural network has no “creation” ability. From an artistic point of view, NST is just another kind of “printing”. In the past, printing was copying images, and now NST was in copying style. If there is no image content to be iterated, it can only produce some meaningless and chaotic pixels. Although some people use image semantic layout technology to make AI create a new picture, for AI, this is just an imitation. There is no emotion in creation, and there may be a big gap between the art we define. But this also presents a whole new problem for artistic creation.
Promising directions for future research on NST mainly focus on two aspects. The first is to solve the current problems of parameter fine-tuning, stroke direction control and neuro-style migration. The second aspect is to derive more extensions from general NST. These interesting extensions can bring benefit to both academia and industry, and may even expand into a brand-new field in the future.

---