Non-convex block-sparse compressed sensing with redundant dictionaries

摘要

Compressed sensing is a novel theory for signal sampling, which breaks through Nyquist/Shannon sampling limitation and makes it into reality that one can efficiently collect and robustly reconstruct a sparse signal. However, some signals exhibit additional structures in some redundant dictionaries, which is called block-sparse signal. In this study, non-convex block-sparse compressed sensing with redundant dictionaries is investigated. Under the block D-RIP condition (√2/2) ≤ δ2k|τ <; 1, a sufficient condition for robust signal reconstruction with redundant dictionaries by mixed ℓ2/ℓp(0 <; p <; 1) minimisation is established. Furthermore, the authors' theoretical results show that, under the assumption that (√2/2) ≤ δ2k|τ <; 1, p ∈ (0,p̂], where p̂ = {1.6835(1-δ2k|τ), δ2k|τ ∈[[√2]/2,0.73) 0.45418, δ2k|τ ∈(0.73,0.7983) 2.2522(1-δ2k|τ), δ2k|τ ∈[0.7983,1), then the block k-sparse signal can be stably reconstructed via non-convex ℓ2/ℓp minimisation with redundant dictionaries in the presence of noise. Particularly, this improves the existed result when the block-sparse signal degenerate to the conventional signal case. Besides, the authors also obtain robust reconstruction condition and error upper bound estimation when the block number is no more than four times the sparsity of the block signal (d ≤ 4k). Moreover, the numerical experiments to some extent testify the performance of non-convex ℓ2/ℓp(0 <; p <; 1) minimisation with redundant dictionaries.