|
Inception residual attention network for remote sensingimage super-resolution B+ Q7 p y4 ]; b7 ~
# u' w3 ~' e# L3 x+ S. t; f: \; [8 \; I6 I: G
How to enhance the spatial resolution for a remote sensing image is
, I( J* m, Q5 [$ M0 l# Kan important issue that we face. Many image super-resolution (SR)
! K* ?. [* t- f! C' Vtechniques have been proposed for this purpose and deep con
1 B$ i/ x# [1 I: _9 ?9 @1 O* |* ?volutional neural network (CNN) is the most effective approach in ' ~$ P( G# @& [% `% R5 L% A
recent years. However, we observe that most CNN-based SR meth
5 l1 ?$ m% U. x8 k t4 ]) Pods treat low-frequency areas and high-frequency areas equally, ) c. z! G2 M9 Q- F' x
hence hindering the recovery of high-frequency information. In this - W8 p+ w% E' K1 y* s9 J: d% j
paper, we propose a network named inception residual attention
: e0 N: n7 J3 C8 J# J9 Z, Snetwork (IRAN) to address this problem. Specifically, we propose
. C; E' F5 p6 }) A/ O4 D. ja spatial attention module to make the network adaptively learn
& R" @/ p9 Z3 U, ` Q" p/ x2 `the importance of different spatial areas, so as to pay more atten d- P3 K: y5 P1 H
tion to the areas with high-frequency information. Furthermore, we # ]1 h9 w' Z, G& Q% G
present an inception module to fuse local multilevel features, so as : k- j6 l( T. s; ^9 F# r/ |
to provide richer information for reconstructing detailed textures. In 8 L8 @5 O3 _: O! {: R( G
order to evaluate the effectiveness of the proposed method, a large 7 y* Z4 Y6 G7 I3 Q
number of experiments are performed on UCMerced-LandUse data " O/ ], Z6 v8 Q2 s2 t
set and the results show that the proposed method is superior to
) h3 i; M& G, t$ \5 k6 Othe current state-of-the-art methods in both visual effects and
0 o& a1 _9 o/ ?& F6 N( Aobjective indicators.
* M/ t# i: V$ |( k* E, ^( e6 j
" v/ k4 Z+ W' t
& N. R+ C/ O' Q0 C- _9 W
( l5 p$ o- X& d! h! l. @! O/ {. x& U
2 c/ } \ y* j& P( g |