Compressive sensing is a technique that allows for sampling and reconstruction of sparse signals from far fewer samples or measurements than traditional methods. It addresses the exponential growth of digital data by enabling better sensing of less data. A sparse signal can be accurately reconstructed from incomplete measurements if it has a sparse representation in some domain, the measurements have a noise-like pattern, and non-linear reconstruction is used. Compressive sensing has been applied to MRI to reduce scan time by sampling the k-space sparsely and reconstructing the image using compressed sensing techniques.
2. Compressive Sensing
S Exponential growth of data
S 48h of video uploaded/min on Youtube
S 571 new websites/min
S 100 Terabytes of dada uploaded on facebook/day
S How to cope with that amount
S Compression
S Better sensing of less data
5. S Shanon’s sampling theorem
S Full recovery under Nyquist sampling frequency?
S Yes if fulfilling 3 criteria
S Sparsity
S Incoherence
S Non linear reconstruction
Compressive Sensing
7. Sparsity
S Desired signal has a sparse representation in some domain D
S x of length N
S x is K sparse
S x has K non zeros components in D
S Can be reconstructed using only M measurments (K<M<N)
Wavelet transform
8. Incoherence
S Random subsampling must show “noise-like” pattern in the
transform domain
S Undersampling introduces noise
S Randomly undersampled Fourier space is incoherent
12. S RIP
1-eK( ) £
Ax 2
2
x 2
2
£ 1+eK( )
Restricted Isometry Property
13. S If Φis a M x N Gaussian matrix
S M > O ( Klog(N))
S If Ψ is a N x N sparsifying basis
S ΦΨ satisfies the RIP condition
Restricted Isometry Property
M.Rudelsonand, R.Vershynin, “On sparse reconstruction from Fourier and Gaussian
measurements,” Commun. Pure Appl. Math., vol. 61, no. 8, pp. 1025–1045, 2008.
14. S Gerhard Richter (1024 colours - 1974)
Restricted Isometry Property
15. Measurements required
S How many measurements required
S M ≥ K+1
S Only if
S No noise
S Real sparse signal
S But
S NP hard problem (exponential numbers of subsets)
27. S Random Fourier matrix satisfies the RIP condition:
S M randomly chosen columns of NxN DFT matrix
S M = O ( K.log (N) )
Direct Fourier Transform
M.Rudelsonand, R.Vershynin, “On sparse reconstruction from Fourier and Gaussian
measurements,” Commun. Pure Appl. Math., vol. 61, no. 8, pp. 1025–1045, 2008.