The integration of model-based and deep network-based beamforming represents a promising solution for sound imaging. Its core logic hinges on leveraging the complementarity of these two approaches: physical models address neural networks’ reliance on large datasets and poor interpretability, while neural networks compensate for physical models’ limitations in adapting to complex scenarios. By unfolding the iterative steps of the fast iterative shrinkage thresholding algorithm (FISTA) into network layers and converting FISTA’s fixed parameters into learnable ones, an enhanced sound imaging method is presented. This method achieves adaptability to diverse scenarios while retaining a certain level of physical interpretability. Simulation results demonstrate its superior performance compared to alternative approaches.
Wu et al. (Wed,) studied this question.