The research team of the Machine Learning Laboratory,

Paper Accepted at the Outstanding International Conference ACM International Conference on Multimedia Retrieval (ICMR) 2023

▲ (Top row, from left) Choi Ha-ram and Na Cheol-woong, master's students, Department of Computer Science & Engineering

(Bottom row, from left) Kim Jin-seop, master's student, interdisciplinary programme in Artificial Intelligence; Professor Yang Ji-hoon, Department of Computer Science & Engineering

A paper written by graduate students of the Machine Learning Laboratory — Choi Ha-ram (Journalism and Broadcasting '16; master's, 4th semester; sole first author) and Na Cheol-woong (Computer Science & Engineering '15; master's, 2nd semester) of the Department of Computer Science & Engineering, and Kim Jin-seop (master's, 2nd semester) of the interdisciplinary programme in Artificial Intelligence — with supervisor Professor Yang Ji-hoon (corresponding author), has been accepted at the international AI conference ACM International Conference on Multimedia Retrieval 2023 (ICMR 2023). ICMR is listed at an adjusted IF of 1 among the outstanding international conferences in computer science under the BK21 Plus programme.

The paper, titled 'Exploration of Lightweight Single Image Denoising with Transformers and Truly Fair Training', introduces baselines for image denoising by appropriately adapting seven recent lightweight transformer models. The effectiveness of transformer models has been demonstrated in much research, but denoising using lightweight transformer models has scarcely been studied to date. The various baselines presented in the paper are therefore expected to become a very important foundation for related research.

The team also pointed out that most prior research on denoising has compared model performance in ways that are not fair. In training denoising models, portions of images are randomly cropped and used as training data. The paper shows that when this randomness is not properly controlled, the resulting performance differences are conspicuous. To overcome this limitation, the paper introduces a genuinely fair method of performance comparison by controlling the randomness that arises during deep learning training.

Finally, the paper presents experimental and valuable analysis of several key considerations for constructing lightweight image denoising deep learning models and improving their performance. The paper was accepted at ICMR 2023 in recognition of this contribution in setting out effective research directions for future researchers.

▲ Denoising performance comparison of the seven lightweight baseline models presented in the paper (SwinIR-light, ELAN-light, NGswin, Restormer-light, Uformer-light, CAT-light, ART-light)

(Figure caption: the goal is to restore the image marked Low Quality to the image marked High Quality. SwinIR (large) is a large, non-lightweight model; notably, the denoising performance of the lightweight baseline models is not greatly inferior to that of the large model as perceived by humans.)

▲ Performance comparison between large denoising models (above: SwinIR, Restormer) and the lightweight denoising models introduced in the paper (below: SwinIR-light, Restormer-light)

(Figure caption: the goal is to restore the image marked Low Quality to the image marked High Quality. The figures below each model name are PSNR (Peak Signal-to-Noise Ratio), a performance measure of image restoration accuracy, and the number of trainable parameters indicating model size. The lightweight models, some 10–20 times smaller than the large models, fall behind numerically but achieve comparable denoising performance in terms of human perception.)

Choi Ha-ram said: "The work I did as an extension of NGswin, my paper recently accepted at CVPR 2023, has been published at the outstanding international conference ICMR 2023, laying the foundation for continuing this line of research. It is a great honour to have papers accepted at outstanding international conferences twice within two months, and I will strive further for the advancement of AI in Korea. Finally, this week is my parents' wedding anniversary, and I hope these successive paper acceptances make a good gift; I am grateful to my family for always supporting me."

ICMR 2023, organized by ACM (Association for Computing Machinery), will be held in Thessaloniki, Greece, from 12 to 15 June.

▶ Go to the paper: https://arxiv.org/abs/2304.01805

▶ Paper summary and code: https://github.com/rami0205/LWDN