ETRI Develops Technology to Edit Facial Photos Without Professional Skills
Backbone Network·SC-FEGAN Technology Drives AI Ecosystem Implementation
Releases Object Recognition Training Data for Urban Environments Needed for AI Learning

Core algorithms for visual artificial intelligence technology, technology for editing facial photos without professional expertise, and image data necessary for visual intelligence learning are being made available to the public.
The Electronics and Telecommunications Research Institute (ETRI) is contributing to the development of the domestic artificial intelligence industry ecosystem by releasing VoVNet, a backbone network which is a core technology necessary for implementing visual artificial intelligence such as object recognition and action tracking, and SC-FEGAN, technology that enables facial editing without Photoshop.

In addition, the institute is disclosing 200,000 object recognition training images targeting 560 types of urban environment objects necessary for training visual artificial intelligence technology.

Unlike the human eye, computers require complex processes for image discrimination and recognition in video; however, the backbone network released by ETRI can locate and extract features of objects in photographs and analyze the information to create models with artificial neural networks.

The backbone network is capable of implementing functions such as object detection, object segmentation recognition by part, and facial recognition, and is evaluated as a core foundational technology for visual intelligence. Through this technology, developers can create added value by implementing desired services or innovative functions.

SC-FEGAN technology is capable of adding accessories that a person in a photograph is not wearing or changing hairstyle, facial expression, and other features. Even if portions are damaged by doodling or blank spaces appear, desired content can be simply drawn and restored. In essence, it is an algorithm capable of various types of editing specialized for facial photographs.

This technology employs GAN (Generative Adversarial Network) technology, one of the deep learning techniques, which artificially creates data and generates realistic-looking data by discriminating it. However, while this technology is effective for synthesizing or transforming images, it has the disadvantage of not reflecting user intent or conditions.

Accordingly, ETRI researchers developed the technology to obtain desired output values by inputting values, and through this approach, when a portrait photograph and input values desired by the user are entered into the algorithm, an image that meets the conditions while naturally harmonizing with the surrounding environment in the photograph can be created.

ETRI expects this technology can significantly reduce work time and improve result quality in fields such as computer graphics, web design, and industrial design.

The additionally disclosed data includes high-quality data essential for visual artificial intelligence technology learning. This data contains objects appearing in urban environments typically recorded by utility poles, traffic signals, and CCTV.

Park Jong-yeol, Director of ETRI's Visual Intelligence Research Laboratory, stated, "While domestic visual artificial intelligence technology is growing rapidly, dependence on overseas technology is gradually increasing. We will actively support the public release of related technologies so that domestic industry, academia, and research institutions can secure more competitive technologies and build an ecosystem."

Meanwhile, ETRI plans to continuously release core foundational technologies related to visual intelligence and high-quality data, while also developing technology to easily edit other objects such as refrigerators and furniture.