Open tkuri opened 2 years ago
Deep Image Prior(DIP)は様々な逆問題に効果的だが、ノイズ除去に対しては(ガウスノイズを仮定したとしても)あまり上手く機能せずBM3Dよりも性能が低い。そこで最適化の進捗を監視する自由度の概念を導入し、ノイズにフィットする前の原理的な停止基準を提案し効果を実証。
https://openaccess.thecvf.com/content/ICCV2021/html/Jo_Rethinking_Deep_Image_Prior_for_Denoising_ICCV_2021_paper.html
https://github.com/gistvision/DIP-denosing
論文概要
Deep Image Prior(DIP)は様々な逆問題に効果的だが、ノイズ除去に対しては(ガウスノイズを仮定したとしても)あまり上手く機能せずBM3Dよりも性能が低い。そこで最適化の進捗を監視する自由度の概念を導入し、ノイズにフィットする前の原理的な停止基準を提案し効果を実証。
https://openaccess.thecvf.com/content/ICCV2021/html/Jo_Rethinking_Deep_Image_Prior_for_Denoising_ICCV_2021_paper.html
Code
https://github.com/gistvision/DIP-denosing