"Deep Photo Style Transfer" which transfers "appearance features of photographs" to another picture using deep learning

Deep learningBy giving a style (visual feature) of another image to the base image by using the "Deep Photo Style Transfer"Is published on GitHub, a shared platform for software development projects.

GitHub - luanfujun / deep - photo - styletransfer: Code and data for paper "Deep Photo Style Transfer"
https://github.com/luanfujun/deep-photo-styletransfer

In the past, it seems that there was an attempt to transfer another image style to the base image using deep learning etc. However, even using pictures in both the base style, there was a problem that the output image had a "picture distortion" that is not like a picture. there,A new study, By limiting the conversion from the input image to the output image only to the local affine in the color space, only elements such as time, weather, and season are transferred from the style image to the base photo, and the photorealistic image It was able to output it.

This result is reflected in "Deep Photo Style Transfer", but what kind of image can actually be generated can be understood by looking at the image below.

Select the landscape photograph of the city where the building stands in the base image.


Next, select the style you want to give to the base photo. If you consider style as odd features, it's ok, for example if you let me learn the style of the following photos ...


We captured the visual features of the second picture "Night view", and the base photo turned into a night view photo.


Style to incorporate into the base picture When another picture is taken at night ......


Night view pictures with different appearance characteristics can be generated.


Besides, based on the mediocre cloudy sky and the scenery of the lake ... ...


Taking in the style of the landscape photograph which is characterized by bright red clouds ......


A red cloud was added to the mediocre landscape photograph, and a fantastic landscape photograph was generated.


Furthermore, based on the yellow dahlia ... ....


If you learn the style from a picture of a crimson rose ... ...


You can generate pictures of deep red dahlia.


Cases where I took pictures of a certain place but I was wondering "What kind of photos can be taken during the autumn leaves season?"


If you set the photo of autumn leaves season to style ......


Even if you have not seen it, you can create a photo of the autumn leaves season at a particular place.


"Deep Photo Style Transfer" is a library for developing open source machine learning and speech recognition applicationsTorch"It was developed based on the open source project on GitHub, and it is only for use for educational / non-commercial use.

GitHub - luanfujun / deep - photo - styletransfer: Code and data for paper "Deep Photo Style Transfer"