Using photo editing software to enhance photos is fairly common these days. Whether we want to lessen the effect of imperfections such as dreaded ‘red-eye’ or even add elements that weren’t there before, photo editing is definitely here to stay.
Some applications such as Instagram also come with filters for photos which alter the look of photos. Some of these filters essentially make it seems as if your photo was taken under different lighting conditions.
Recently however, a team of researchers from Cornell University consisting of professor Kavita Bala and doctoral student Fujun Luan in collaboration with Sylvian Paris Eli Shechtman from Adobe have created a software which can transfer the style of one image to another image.
The objects in the photographs are relatively unchanged, with the style transfer technique focused mainly on altering colors.
This technique involves, what is essentially a donor image from which the desired style will be copied. The style is copied to the desired image while maintaining the same structure and the final result is quite impressive.
The final result does not look out of place at all. The objects in the photographs are relatively unchanged, with the style transfer technique focused mainly on altering colors.
Please see the example below of the results of the software.
How It Works
Essentially, this solution extracts the features of the donor image and infuses them with a target image
The team formulated a clever deep learning solution which employed a neural network layer in order to execute the technique.
Deep learning, as the name suggests, is a method of computer learning. It is capable of completing learning tasks using neural networks. A neural network is a computer system capable of learning based on data it surveys. This system is based on the biological configuration of a brain.
Initially, it was difficult for the team to produce photos which you could tell came from a certain donor image. They came up with a clever solution which alters the photo while maintaining the boundaries and edges of the original image.
Essentially, this solution extracts the features of the donor image and infuses them with a target image.
Other solutions exist which can transfer styles between images but they tend to resemble paintings, even if the ‘donor’ image is a photo.
The Cornell/Adobe team’s solution impressively manages to maintain its photorealism with very little distortion.
This software could be used to transfer certain characteristics such as time of day and weather for instance.
This technique can be used for a number of purposes. Of course, it just looks cool. However the technique can be used to transfer a variety of characteristics from an existing photo which the user of such a software might desire to have in another photo.
This software could be used to transfer certain characteristics such as time of day and weather for instance. This technique can also be used to add artistic edits to photos, based on the donor image.
This software could prove quite the powerful image editing tool. The end user would be able to quickly give their photos a style make over with virtually no manual labor. All that would be required to create a masterpiece would be a suitable donor image.