Noise-aware physics-consistent neural deconvolution for active sonar beamforming maps

High-frequency shallow-water active sonar commonly forms angle–range beam maps by matched filtering and delay-and-sum beamforming, but the resulting conventional beamforming outputs often suffer from limited resolution, strong sidelobes, and severe performance degradation under low signal-to-noise ratio (SNR) and coherent interference. Beam-domain deconvolution methods model the beam map as a sparse reflectivity distribution blurred by a point-spread function (PSF), yet their inversion is ill-posed, sensitive to noise, and further challenged by shift-variant PSFs in wide-field angle–range imaging. This work proposes a noise-aware deconvolution beamforming convolutional neural network (NA-DBF-CNN) that couples a fully convolutional encoder–embedding–decoder backbone with an explicit physics-guided consistency constraint. The network is trained using a hybrid objective consisting of a peak-emphasized supervision loss and a SNR-weighted beam-domain data-consistency loss derived from the measurement model, where the noise-aware weight reflects sample-dependent reliability under mixed-SNR training. Monte Carlo simulations and lake experiments confirm that by enforcing consistency with the matched-filtering delay-and-sum forward projection, NA-DBF-CNN alleviates the reliance on shift-invariant PSF approximations and improves robustness in challenging multi-target scenarios.