ABSTRACT Conventional active sonar processing chains, which typically involve separate stages of beamformer and matched filter, are often suboptimal and struggle to adapt to complex underwater environments. This paper proposes a novel end‐to‐end deep learning framework that directly maps raw multichannel frequency‐domain data to azimuthal detection results, bypassing the need for traditional processing modules. The input to the network is a two‐dimensional feature map derived from the complex frequency‐domain data of a horizontal line array, with dimensions representing the number of hydrophones and frequency bins. The output is a vector representing the likelihood of target presence across an azimuthal sector. The backbone of the network consists of specially designed convolutional layers and attention layers, which are designed to meet the requirements of active sonar detection tasks. Experimental results from both simulation and lake trials demonstrate that the proposed end‐to‐end model demonstrates improved detection performance compared to conventional methods, achieving higher detection probability whilst maintaining a lower false alarm rate in challenging environments.
Qing et al. (Thu,) studied this question.