-
Advertorial
-
FOCUS
-
Guide
-
Lifestyle
-
Tech and Vogue
-
TechandScience
-
CHTF Special
-
Nanshan
-
Futian Today
-
Hit Bravo
-
Special Report
-
Junior Journalist Program
-
World Economy
-
Opinion
-
Diversions
-
Hotels
-
Movies
-
People
-
Person of the week
-
Weekend
-
Photo Highlights
-
Currency Focus
-
Kaleidoscope
-
Tech and Science
-
News Picks
-
Yes Teens
-
Budding Writers
-
Fun
-
Campus
-
Glamour
-
News
-
Digital Paper
-
Food drink
-
Majors_Forum
-
Speak Shenzhen
-
Shopping
-
Business_Markets
-
Restaurants
-
Travel
-
Investment
-
Hotels
-
Yearend Review
-
World
-
Sports
-
Entertainment
-
QINGDAO TODAY
-
In depth
-
Leisure Highlights
-
Markets
-
Business
-
Culture
-
China
-
Shenzhen
-
Important news
在线翻译:
szdaily -> Tech and Science -> 
AI can produce celebrity-like face images
    2018-03-07  08:53    Shenzhen Daily

虚拟人脸照片, 全都逼真到家

They may look like they came straight from the cover of a magazine, but these images show “famous” people who don’t actually exist.

Scientists have combined a pair of artificial intelligence (AI*) systems to generate photo-realistic faces of fake celebrities. In a video showing the process, faces can be seen shifting and warping* as the networks gradually tweak each feature over the course of 20 days.

The system, developed by researchers at U.S. microchip maker Nvidia, uses a new type of algorithm* called a “generative adversarial network” (GAN).

Artificial neural networks are systems designed to think like the human brain, and in a GAN, two of the networks are pitted against one another. One network is tasked with generating something, like an image or audio pattern, while the other plays an adversarial role by challenging the results of the first. The idea is that the two create sharper results by competing against one another.

Nvidia, based in Santa Clara, California, trained its software using images from a publicly available database of photos of famous people.

The generative network of the two would then create a low-resolution image based on this large dataset. The discriminator* network would then assess the work of the first by comparing them to the real-world images.

As the system improved the researchers added new layers that looked at higher-resolution details until the GAN ended up with photo-realistic images.(SD-Agencies)

深圳报业集团版权所有, 未经授权禁止复制; Copyright 2010, All Rights Reserved.
Shenzhen Daily E-mail:szdaily@szszd.com.cn