Self-attention mechanism for distributed compressive sensing

This letter addresses the joint sparse recovery problem with multiple measurement vectors in compressive sensing (CS). In this letter, it is assumed that the sparse vectors are related to each other but the specific dependencies are unknown. The authors propose a data-driven approach that relies on a self-attention mechanism to learn the sparse structure within and between sparse vectors automatically. A greedy method is then used to reconstruct the sparse vectors. The numerical experiments conducted on two real-world datasets show that the proposed method outperforms the model-based Bayesian methods and data-driven method based on long short-term memory.