2018年8月1日
Iterative applications of image completion with CNN-based failure detection
Journal of Visual Communication and Image Representation
- ,
- ,
- ,
- ,
- 巻
- 55
- 号
- 開始ページ
- 56
- 終了ページ
- 66
- 記述言語
- 英語
- 掲載種別
- 研究論文(学術雑誌)
- DOI
- 10.1016/j.jvcir.2018.05.015
- 出版者・発行元
- Academic Press Inc.
Image completion is a technique to fill missing regions in a damaged or redacted image. A patch-based approach is one of major approaches, which solves an optimization problem that involves pixel values in missing regions and similar image patch search. One major problem of this approach is that it sometimes duplicates implausible texture in the image or overly smooths down a missing region when the algorithm cannot find better patches. As a practical remedy, the user may provide an interaction to identify such regions and re-apply image completion iteratively until she/he gets a desirable result. In this work, inspired by this idea, we propose a framework of human-in-the-loop style image completion with automatic failure detection using a deep neural network instead of human interaction. Our neural network takes small patches extracted from multiple feature maps obtained from the completion process as input for the automated interaction process, which is iterated several times. We experimentally show that our neural network outperforms a conventional linear support vector machine. Our subjective evaluation demonstrates that our method drastically improves the visual quality of resulting images compared to non-iterative application.
- ID情報
-
- DOI : 10.1016/j.jvcir.2018.05.015
- ISSN : 1095-9076
- ISSN : 1047-3203
- SCOPUS ID : 85047619410