Improving Neural Style Transfer by Incorporating Mid-Level Representation
Jordan Raji Lazo Cahua, José Eduardo Ochoa-Luna · 2021
Gatys et al. [1] has demonstrated that convolutional neural networks (CNN) are able to separate the style and content of images and then recombine them into an artistic image. This process is known as neural style transfer (NST) and has had repercussions in both academia and industry. This paper covers the approach based on image optimization that allows the selection of artistic styles arbitrarily; however, the process has a high computational cost. For this reason, it is proposed to make some modifications that allow performing the neural style transfer with a lower computational cost, in addition to adding a similarity indicator that determines whether the synthesized image has reached the desired style using mid-level representation techniques. These modifications have given better visual qualitative results, in addition, the execution time of an iteration has improved up to 1.53× times compared to other similar methods, and with the similarity indicator, it has been possible to reduce the number of iterations necessary to obtain the synthesized image.