Notice

One BMIPL’s paper, “Extending Stein’s unbiased risk estimator to train deep denoisers with correlated pairs of noisy images,” by Magauiya Zhussip, Shakarim Soltanayev and Se Young Chun has been accepted to NeurIPS 2019 (acceptance rate: 21.2%). Magauiya Zhussip and Shakarim Soltanayev just finished their MS program at BMIPL, UNIST. NeurIPS is one of the top machine learning conferences. Last year, there have been a few works on training deep learning based denoisers without ground truth by NVIDIA (noise2noise) and BMIPL@UNIST (SURE based training). Our work extended conventional Stein’s unbiased risk estimator (SURE) to deal with correlated pairs of noisy data and applied it to training deep denoisers. Our work also addressed issues in imperfect ground truth for training deep denoisers.