▲ (From left) Choi Ha-ram (master's student) and Yang Ji-hoon (supervisor)

The paper 'N-Gram in Swin Transformers for Efficient Lightweight Super-Resolution', written by graduate student Choi Ha-ram of the Department of Computer Science & Engineering (master's, 4th semester, first author) and Professor Yang Ji-hoon (corresponding author), has been accepted at Computer Vision and Pattern Recognition (CVPR) 2023, the world's leading international conference in artificial intelligence.

The paper develops NGswin, a lightweight and efficient model for super-resolution technology that restores low-resolution images to high resolution. NGswin is the first case of introducing the N-Gram methodology, chiefly used in text analysis, into image restoration with efficient and competitive performance. The paper also applies the N-Gram methodology to the prior model SwinIR-light to improve its performance, developing SwinIR-NG and achieving world-leading performance in lightweight image super-resolution.

▲ The architecture of NGswin, an efficient model for image super-resolution (a, b)

The N-Gram methodology, proposed for image restoration for the first time in the world (c)

The paper was accepted at CVPR 2023 in recognition of this academic contribution. CVPR 2023, organized by IEEE/CVF, will be held at the Vancouver Convention Centre in Canada from 18 to 22 June.

Choi Ha-ram, who majored in journalism and broadcasting and in Big Data Science as an undergraduate at the University, said she hopes that having a CVPR paper accepted during her master's studies will be a great motivation to the University's graduate students and to students from non-technical backgrounds.