Abstract
This paper presents a two-stream learning-based compressive sensing network with a high-frequency compensation module (TSLCSNet) that betters restores the detailed components of an image during the image denoising process. The proposed two-stream network consists of a compressive sensing network (CSN) and a high-frequency compensation network (HCN). CSN restores the main structure of the image, while HCN adds the detail that is not obtainable from the CSN. To improve the performance of the proposed model, we add an incoherence loss function to the total loss function. We also employ an octave convolution to allow the two-stream network to communicate in order to extract less redundant and more compressive features. Representative experimental results show the superiority of the proposed TSLCSNet and TSLCSNet+ compared to state-of-the-art methods for the removal of synthetic and real noise.
Original language | English |
---|---|
Article number | 9464234 |
Pages (from-to) | 91974-91982 |
Number of pages | 9 |
Journal | IEEE Access |
Volume | 9 |
DOIs | |
Publication status | Published - 2021 |
Bibliographical note
Funding Information:This work was supported by the Agency for Defense Development of Korea under Grant UD190005DD.
Publisher Copyright:
© 2013 IEEE.
Keywords
- ISTA
- compressive sensing
- deep learning
- denoising
ASJC Scopus subject areas
- General Computer Science
- General Materials Science
- General Engineering