Naslov
Multilevel Subsampling of Principal Component Projections for Adaptive Compressive Sensing

Autori
Vlašić, Tin ; Seršić, Damir

Vrsta, podvrsta i kategorija rada
Radovi u zbornicima skupova, cjeloviti rad (in extenso), znanstveni

Izvornik
Proceedings of the 12th International Symposium on Image and Signal Processing and Analysis (ISPA 2021) / Lončarić, Sven ; Petković, Tomislav ; Petrinović, Davor - Zagreb : Sveučilište u Zagrebu, 2021, 29-35

Skup
12th International Symposium on Image and Signal Processing and Analysis (ISPA 2021)

Mjesto i datum
Zagreb, Hrvatska, 13.09.2021. - 15.09.2021

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
compressive sensing ; inverse problem ; principal component analysis ; sampling theory ; sparse signal recovery

Sažetak
This paper examines the performance of principal- component-analysis (PCA) projections in compressive sensing (CS). Observed signals are assumed to follow a Gaussian distribution and have the asymptotic sparsity property in a wavelet transform domain. In order to exploit these signal priors, we propose multilevel subsampling of PCA projections in addition to sparsity-promoting $\ell_1$ regularization. The PCA projections are subsampled in levels that correspond to different wavelet scales. The proposed method outperforms universal random projections of standard CS for noise-corrupted measurement setups and compressible signals. Experimental results from simulations conducted on images from the MNIST dataset prove the framework's robustness and good reconstruction ability.

Izvorni jezik
Engleski

Znanstvena područja
Elektrotehnika

POVEZANOST RADA