Gwangju University of Science and Technology Researchers Develop New Method to Remove Image Noise
This approach involves a post-correction network optimized by a self-supervised machine learning framework to improve the quality of unfamiliar images.
Gwangju, Korea, September 20, 2022 /PRNewswire/ — High quality visual displays rendered using the “path tracing” algorithm are often noisy. Modern supervised learning-based denoising algorithms rely on external training datasets, are slow to train, and do not perform well when training and test images are different.Currently a researcher at Gwangju University of Science and Technology, VinAI Research and University of Waterloo We presented a new self-supervised post-correction network that improves denoising performance without relying on references.
Widely used in games, illustrations, and visualizations, high-quality computer graphics are considered the cutting edge of visual display technology. The method used to render high quality, realistic images is known as “path tracing”. monte carlo (MC) A denoising approach based on supervised machine learning. In this learning framework, a machine learning model is first pre-trained on pairs of noisy and clean images and then applied to real noisy images (test images) to be rendered. Although considered to be the best method in terms of image quality, this method may not work well if the test images differ significantly from the images used for training.
A group of researchers, including PhDs, to address this issue.student Jung Hee Buck Associate Professor Bakushougetsu Research Fellow from Gwangju University of Science and Technology, South Korea Vinh Song Hua From VinAI Research at Vietnam,Associate Professor Toshiya Hachisuka from University of Waterloo of Canada, in a new study, proposed a new reference-independent MC denoising method.Their research is now available online July 24, 2022 When was announced in ACM SIGGRAPH 2022 Conference Minutes.
“Existing methods not only fail when the test and training datasets are very different, but also take time to prepare the training dataset for pre-training the network. It is a neural network that can be trained on the fly using only “for pre-training” Dr. Moon explains the motivation for their research.
To achieve this, the team developed a new post-correction approach for denoised images, consisting of a self-supervised machine learning framework and a post-correction network, which is essentially a convolutional neural network for image processing. I suggested. The modified network did not rely on pre-trained networks and could be optimized using the concept of self-supervised learning without relying on references. Moreover, self-supervised models complemented and enhanced traditional supervised models for denoising.
To test the effectiveness of the proposed network, the team applied the approach to existing state-of-the-art denoising methods. The proposed model showed a 3x improvement in rendered image quality compared to the input image by preserving finer details. Moreover, the entire process of on-the-fly training and final inference took just 12 seconds.
“Our approach is the first one that does not rely on pre-training with external datasets. It actually reduces production time and improves the quality of offline rendering-based content such as animations and movies.” To do.” Dr. Moon speculates about potential applications of their research.
In fact, it may not be long before this technology is used for high-quality graphics rendering in video games, augmented reality, virtual reality, and the metaverse.
Original paper title: Self-supervised Post-Correction for Monte Carlo Denoising
journal: ACM SIGGRAPH 2022 Conference Minutes
About Gwangju Institute of Science and Technology (GIST)
82 62 715 6253
SOURCE Gwangju Institute of Science and Technology