Share Email Print

Proceedings Paper

Transfer learning of deep CNN de-noiser prior for Chinese ancient calligraphy tablet image denoising
Author(s): Feihang Ge; Lifeng He; Yuyan Chao
Format Member Price Non-Member Price
PDF $17.00 $21.00

Paper Abstract

Tablet images are significance vehicles for ancient culture heritage. However, due to natural or artificial destruction, there usually exists a large amounts of noises or scratches in the ancient tablet images, and this makes the recognition of interesting objects carved in the ancient very difficult. To deal with this problem, a method based on transfer learning of DnCNN De-noiser Prior was proposed in this paper. Firstly all parameters of all layers of a DnCNN pre-trained in natural images are transferred to our target networks. The initial trained CNN filter weights were then fine tuned with noised Chinese tablet calligraphy images by back-propagation so that they better reflected the noise modalities of tablet image, where Chinese tablet calligraphy structures are concerned to remove isolated small scratches by combing the connected region technique with DnCNN transfer denoising. Experiments on real noised tablet images demonstrate that the proposed method is effective both in image noise removal and image detail preserve compared with existing image denoising methods.

Paper Details

Date Published: 27 November 2019
PDF: 10 pages
Proc. SPIE 11321, 2019 International Conference on Image and Video Processing, and Artificial Intelligence, 1132110 (27 November 2019); doi: 10.1117/12.2550357
Show Author Affiliations
Feihang Ge, Aichi Prefectural Univ. (Japan)
Lifeng He, Aichi Prefectural Univ. (Japan)
Yuyan Chao, Nagoya Sangyo Univ. (Japan)

Published in SPIE Proceedings Vol. 11321:
2019 International Conference on Image and Video Processing, and Artificial Intelligence
Ruidan Su, Editor(s)

© SPIE. Terms of Use
Back to Top
Sign in to read the full article
Create a free SPIE account to get access to
premium articles and original research
Forgot your username?