Abstract: Most existing compressed image sensing algorithms with outstanding reconstruction performance are based on nonlocal self-similarity of natural images, for the reason that reconstruction based on similar-block groups is more prone to exploit sparsity than reconstruction merely based on individual blocks. But at low sampling rates, limited by poor quality of initial reconstruction results, accuracy of similar-block grouping is badly affected by massive reconstructed interference information, leading to an unsatisfying reconstruction result. To solve this problem, a low-rank enhancement reconstruction algorithm based on GSR is proposed. Firstly, a hybrid filtering reconstruction method is proposed in initial reconstruction that BM3D filtering is introduced to attain a better initially reconstructed result. Then a low-rank enhancement pretreatment is conducted before similar-block grouping to help grouping operation focus more on the key features. Simulation results indicate that compared with GSR, the proposed algorithm possesses a better reconstruction performance at low sampling rates.

Key words: compressed image sensing, low sampling rates, hybrid filtering, key features, lowrank enhancement