ETRI researchers develop core algorithms of AI-based visual technology

The Electronics and Telecommunications Research Institute (ETRI) announced Friday that core algorithms of artificial intelligence-based visual technology have been developed by visual intelligence researchers of its AI research center. It will release 200,000 images of 560 city objects that can be used to train visual AI along with the algorithms that analyze the information extracted from the pictures. The new algorithms and data set are expected to help the Korean AI industry develop new services and functions.

The research team uploaded the algorithms, codes and a manual on Github through VoV Net where AI algorithms can be easily shared with developers. It also shared pictures for AI training, which include photographs of utility poles, traffic lights and cars captured by CCTV cameras across cities.

The researchers also released a new AI-based technology that can naturally edit a human face in pictures without using a professional editing program such as Adobe Photoshop. The new technology allows users to easily add an accessory or touch up hair and facial expressions. It is basically a face-editing algorithm that offer various features.

This technology utilizes the deep learning technique GAN, which is good at creating fake images that look realistic but cannot fully apply the user’s intention or requests. Combined with the GAN, the newly-released technology helps the user edit the image the way they want.

It is also incredibly easy to use – all it needs to create the image is a picture and input from the user, which it then combines together without making it look unnatural.

“Korean visual artificial intelligence is developing at a rapid pace but relies heavily on overseas technologies,” said senior researcher Park Jong-yeol at the institute. “We will spare no effort to support industry, academia and research institutions so that various environments can be created for visual AI technology development.


reborn@donga.com