Spectral Line Enhancement for Noise-Resilient Passive Sonar Detection Using Dual-Attention-Guided Wavelet Domain FISTA-Net

Robust detection of underwater targets by passive sonar is hindered by weak tonal emissions masked by strong ambient noise in the marine environment. These narrowband components, represented as spectral lines in the LOFAR spectrum of received acoustic signals, are critical for target recognition but remain difficult to enhance under low signal-to-noise ratio (SNR) conditions. We propose a spectral enhancement framework that combines a frequency domain adaptive line enhancer (ALE) with a dual-attention guided wavelet domain fast iterative shrinkage thresholding algorithm network (FISTA-Net). The ALE stage suppresses broadband noise while preserving spectral structures, whereas wavelet domain FISTA employs multiresolution analysis and attention-guided weighting to enhance weak spectral features under sparsity constraints. Experimental results show an average SNR improvement of 4.87 dB over block-processed sparsity- based on ALE and 6.37 dB over conventional ALE. To further study the recognition accuracy of the proposed method, a common ResNet based classifier is used. The proposed method outperforms existing approaches, achieving higher classification accuracy and demonstrating its effectiveness for underwater target detection and recognition.