講演抄録/キーワード |
講演名 |
2021-08-26 10:00
ノイズあり複数画像の非剛体位置合わせ ○浅海標徳・西村和也・ソン ホン・林田純弥(九大)・関口博之・八木隆行(Luxonus)・佐藤いまり(NII)・備瀬竜馬(九大) PRMU2021-7 |
抄録 |
(和) |
(まだ登録されていません) |
(英) |
We propose a deep non-rigid alignment network that can simultaneously perform non-rigid alignment and noise decomposition of images despite severe noise and sparse errors. To address this challenging task, we introduce a low-rank loss in deep learning under the assumption that a batch of well-aligned and well-denoised images should be linearly correlated, and thus the matrix consisting the images should be low-lank. It allows us to remove the noise and corruption from input images in a self-supervised learning manner (i.e., t does not require any supervised data). In addition, we introduce a self-attention technique in order to aggregate the information about corruption from the batch of images. To the best of our knowledge, it is first attempt to introduce a low-rank loss for unsupervised deep alignment. Experiments using toy data and real medical image data demonstrate the effectiveness of the proposed method. |
キーワード |
(和) |
教師なし / 深層学習 / 位置合わせ / 欠損補完 / / / / |
(英) |
Unsupervised / Deep Learning / Alignment / Sparse complement / / / / |
文献情報 |
信学技報, vol. 121, no. 155, PRMU2021-7, pp. 1-6, 2021年8月. |
資料番号 |
PRMU2021-7 |
発行日 |
2021-08-19 (PRMU) |
ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
PRMU2021-7 |