Inception residual attention network for remote sensing image super-resolution

杨利霞

Inception residual attention network for remote sensingimage super-resolution

& R7 d" a5 }# c6 ]' V: w
( f4 K# L6 a: V: S) q; T2 i- f! x
/ f8 `2 q* n$ {. HHow to enhance the spatial resolution for a remote sensing image is
an important issue that we face. Many image super-resolution (SR)
. A: _, J+ f9 ?$ A( Y: T+ }2 ]  w( mtechniques have been proposed for this purpose and deep con
8 e  T9 x! ]1 {* k7 t% avolutional neural network (CNN) is the most effective approach in
! Y5 J6 i- |( u- o: {6 }' f8 Orecent years. However, we observe that most CNN-based SR meth
: c, j4 r5 i; m$ ^" @5 Bods treat low-frequency areas and high-frequency areas equally,
' s0 ^+ K7 g9 mhence hindering the recovery of high-frequency information. In this
4 y. P7 _  H8 O* A8 mpaper, we propose a network named inception residual attention
9 @" F3 U, P# i$ D2 a- Knetwork (IRAN) to address this problem. Specifically, we propose
  O6 Z8 F7 N- ta spatial attention module to make the network adaptively learn
* K/ {, Z: b/ T+ `2 B/ }the importance of different spatial areas, so as to pay more atten
) @: d3 V: J" q: Qtion to the areas with high-frequency information. Furthermore, we
+ l, T9 k+ G6 Hpresent an inception module to fuse local multilevel features, so as
& `: V- h5 N( M+ f$ \to provide richer information for reconstructing detailed textures. In
7 v0 N6 O2 Z! k- t9 qorder to evaluate the effectiveness of the proposed method, a large
9 Z7 @/ K6 N' h9 l: U' ]number of experiments are performed on UCMerced-LandUse data
) \" X- m6 h5 O# d" ]; B9 Tset and the results show that the proposed method is superior to
2 j1 y1 \0 L. T  t4 I) Sthe current state-of-the-art methods in both visual effects and
9 o& K; h5 Q: L+ a, S+ f4 J+ hobjective indicators.

6 y) y) h7 a3 s- ^9 R

' t. Y( x( a) K$ L4 X) u0 D
- S" s* o- g& k3 x