Transferred Multi-Perception Attention Networks for
/ T1 K/ \2 l9 r/ T4 S' oRemote Sensing Image Super-Resolution
2 {8 Z% `+ [! e6 I% i4 @1 d& C" F. _! N @# `
& T/ F( u- v. |+ v$ [$ @0 E6 i5 T- O) Q0 A# }4 v& Q
Image super-resolution (SR) reconstruction plays a key role in coping with the increasing# I: m3 [; |/ S/ Z t1 k9 T
demand on remote sensing imaging applications with high spatial resolution requirements. Though* v0 X) W0 u! B' L) B: U
many SR methods have been proposed over the last few years, further research is needed to improve1 _. x7 V! @9 M# y Z! Q( R6 v
SR processes with regard to the complex spatial distribution of the remote sensing images and the
, z$ R$ a: S, J# o5 C# v$ Odiverse spatial scales of ground objects. In this paper, a novel multi-perception attention network& b' E7 v) j$ ?- A& U
(MPSR) is developed with performance exceeding those of many existing state-of-the-art models.3 C4 s. o( l( ^6 D6 J3 B- W
By incorporating the proposed enhanced residual block (ERB) and residual channel attention group
% J N; y1 S! l6 w! m |(RCAG), MPSR can super-resolve low-resolution remote sensing images via multi-perception learning
; a2 J, R3 z5 I- @0 F2 ~and multi-level information adaptive weighted fusion. Moreover, a pre-train and transfer learning
4 z8 z- f. m0 A3 s l+ w# Jstrategy is introduced, which improved the SR performance and stabilized the training procedure.
y0 Z9 N6 g% X+ _Experimental comparisons are conducted using 13 state-of-the-art methods over a remote sensing' x0 O) R! Y; L# m& i1 i
dataset and benchmark natural image sets. The proposed model proved its excellence in both objective8 x' Q! ]& y+ S4 Q7 E
criterion and subjective perspective.8 f& N8 y+ _5 Q0 t
# u8 V7 I0 j9 o& {0 `9 {0 M7 ~. j4 |( [
|