Abstract
The recovery of sparse signals from their linear mapping on a lower-dimensional space is traditionally performed by finding the sparsest solution compatible with such solutions. This task can be partitioned in two phases: support estimation and coefficient estimation. We propose to perform the former with a deep neural network jointly trained with the encoder that divines a support that is used in the latter phase to estimate the coefficients by pseudo-inversion. Numerical evidence demonstrates that the proposed encoder-decoder architecture outperforms state-of-the-art Compressed Sensing (CS) approaches in the recovery of synthetic ECG signals for a compression ratio higher than 2.5. Further tests on real ECG prove the applicability in real-world scenarios.