A CCTV control room in Seocho District, Seoul. (image: Seocho District Office)
SEOUL, Oct. 25 (Korea Bizwire) — A South Korean research team has developed a new technology that uses artificial intelligence (AI) deep learning to clearly recognize and identify the key features of human faces such as eyes, nose and lips, even in low-resolution images.
The Gwangju Institute of Science and Technology (GIST) said Monday that its researchers had developed a new method of improving facial recognition function by using an “attention map” that delivers the information learned through high-resolution face images to a low-resolution facial recognition model.
The attention map refers to the information that visualizes which areas in input images have a significant impact on the recognition result when a deep learning network carries out the recognition task.
For low-resolution images, facial recognition function tends to sharply deteriorate, making it difficult for investigation agencies and security companies to verify the identification of the subjects.
The research team added a loss function to make the attention map of a low-resolution image learn to be similar to the attention map of a high-resolution image, while making it focus on the areas with a significant impact on the recognition result.
This method contributed to achieving a high recognition accuracy of 47.91 percent, 5 percent higher than the previous best record of 45.49 percent.
In addition to facial recognition, this method also proved effective in the task of sorting out objects in low-resolution images.
Kevin Lee (kevinlee@koreabizwire.com)
