Machine Learning and Knowledge Extraction by Unknown

Author:Unknown
Language: eng
Format: epub
ISBN: 9783030573218
Publisher: Springer International Publishing

(5)

where represents the parameter set of NSAN. We choose Adam algorithm to optimize the loss function.

Fig. 1.Framework of the proposed non-local second-order attention network (NSAN).

Fig. 2.Framework of the proposed non-local second-order attention module (NSA).

3.2 High-Order Enhanced Group (HEG)

We now describe our edge enhanced (HEG) (see Fig. 1), which can be divided into the main branch and the edge enhanced branch. The main branch consists of two region-level non-local (RL-NL) modules [6] and G non-local residual channel attention groups (NRCAG) structure. The RL-NL can capture the long-range information. Each NRCAG further contains M simplified residual channel blocks with local skip connection, followed by a non-local channel attention (NCA) module to exploit feature interdependencies. The edge enhanced branch consists of the padding module and V NRCAG, which can make full use of the edge information and use edge information to enhance channel feature attention.

Stacking residual blocks has been verified that is helpful way to form a deep network in [6, 22, 23]. Nevertheless, deeper network built in such way would lead in performance bottleneck and training difficulty during the problem of gradient vanishing and exploding in deep network. It is known simply stacking repeated block may not to obtain better performance. In order to address this issue, we introduce the NHAG to not only to bypass abundant low-frequency information from LR images, but also facilitate the training of our deep network. Then a HEG in the g-th group is represented as:

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.