Multi-stage iterative compressed sensing framework based on DIFF transformer and ISTA for remote sensing images

In remote sensing applications, image storage and compression require large amounts of memory and consume a lot of battery power. Compressed sensing (CS) can alleviate these problems while shifting the complexity to the decoding end. However, existing methods still face limitations, such as attentio...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Scientific reports Jg. 15; H. 1; S. 38521 - 15
Hauptverfasser: Wang, Rongrong, Fang, Yong
Format: Journal Article
Sprache:Englisch
Veröffentlicht: London Nature Publishing Group UK 04.11.2025
Nature Publishing Group
Nature Portfolio
Schlagworte:
704/2151
704/4111
Algorithms
Compressed sensing
Compression
Convex analysis
Deep learning
DIFF transformer
Humanities and Social Sciences
Image reconstruction
ISTA
Methods
multidisciplinary
Optimization
Remote sensing
Sampling
Science
Science (multidisciplinary)
Signal processing
DIFF transformer
ISTA
Compressed sensing
Image reconstruction
Remote sensing
ISSN:2045-2322, 2045-2322
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In remote sensing applications, image storage and compression require large amounts of memory and consume a lot of battery power. Compressed sensing (CS) can alleviate these problems while shifting the complexity to the decoding end. However, existing methods still face limitations, such as attention noise in standard Transformer architectures and gradient instability in deep-unfolded ISTA variants. To address these issues, we propose a multi-stage iterative compression sensing framework based on DIFF transformer and iterative shrinkage threshold algorithm (ISTA) for remote sensing images, which integrates local details and global structures and pays more attention to context-related information. Our framework includes sampling module and reconstruction module. The sensing matrix of the sampling module is obtained through data-driven training, and the initial reconstruction is performed based on the learned structural information. The reconstruction module uses DIFF Transformer and ISTA stacking and cross-talk to achieve multi-stage iterative reconstruction. Compared with other advanced methods, especially on the NWPU VHR-10 test set, our proposed method is improved by 2.98 dB (14.16 %), 3.15 dB (13.06 %), 2.84 dB (10.30 %) and 0.30 dB (0.01 %) over ISTA-Net+ method. In addition, the proposed method is almost unaffected by the added noise at low sampling rate .
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
content type line 23
ISSN:2045-2322
2045-2322
DOI:10.1038/s41598-025-16727-6