A Single Network for Reconstruction with Dynamic Measurement Rates

In the last decade, deep learning (DL) based signal reconstruction models surpassed traditional compressive sensing (CS) algorithms in quality and speed. A serious downside of DL-based signal reconstruction models is their extreme dependency on a particular measurement pattern and rate used in their training, rendering them inadaptive to changes in measurement rate. This study proposes a joint measurement learning-image reconstruction DL framework designed to process multiple measurement rates. This architecture is based on hierarchical reconstruction with a multi-stage loss function. The evaluations for the proposed model include the effect of the hyperparameters and comparison with single-rate models. Simulation results show that the proposed multi-rate model provides comparable reconstruction performance to single-rate models dedicated to a particular measurement rate and surpasses the alternative solutions for multi-rate reconstruction in image quality.