This “travel camera” is a replacement for old pictures in a new format

How can I get a Lincoln headshot on an iPhone? You should also try this ‘travel camera’ in addition to going back to the 19th century. Countless historical figures have left black and white images for more than 100 years since the advent of photography. These historical images are, however, steadily blurred due to different ages, different technology, and the deterioration of the years. With the look of new cameras, will they all be returned to their original appearance?

Today, in these photographs, the new paper Time-Travel Rephotography from the University of Washington, UC Berkeley, and Google reveals the historical figures on the phone.

Madame Curie, for instance, Edison, Lin Huiyin…

How to fix old pictures with expanded weaknesses

Nothing modern is the restoration of human faces from black-and-white images and even portraits. The recent revival of terracotta warriors, for example, and videos of ancient emperors on social networks have become popular.

So is the recovered previous picture more real?

It’s not the case here. At the end of the year, the defects in Kafka’s face were caused by film technology. You will not thoroughly describe the features and ageing phase of the old film if you merely fill in black and white images.

The author uses the StyleGAN2 framework to map old photos into the space of modern high-resolution photos, so as to achieve all these effects in a unified framework.

The author uses other color sample images as references, which have similar facial features to black and white images, but contain high-frequency details as well as natural colors and lighting. In order to further reduce the perceived identity gap between the input image and modern portraits, they also designed a reconstruction loss especially suitable for old photos.

The structure of the model is as follows:

1. Same level encoder

The low-resolution gray-scale reference image I is used as input to generate the same-level image Is with real colors and similar facial features.

2. Same level color and detail conversion

In order to further limit the color and skin details to match the image of the same level, the model introduces the color conversion loss Lcolor to force the ToRGB layer output of StyleGAN2 to be similar to the image of the same level.

3. Loss of reconstruction of old photos

In order to further reduce the perceived identity gap between the input image and modern portraits, the author specially designed the reconstruction loss Lrecon suitable for old photos, which is robust to the defects of old photos.

4. Potential coding optimization

Superimpose all the previous loss functions to optimize and generate the final image.

According to the author, generating such a 1024 x 1024 photo requires 10 minutes to run on NVIDIA Titan X.

Compared with several other methods of restoring old photos, the effect of “through camera” obviously has a more realistic look and feel.

However, there are still some shortcomings in the “travel camera”.

The first is that there are errors in the restoration of the character distribution and accessories, or details will be lost.

References:

dyTech

cbeta