Noise-aware physics-consistent neural deconvolution for active sonar beamforming mapsa).

Area of Science:

• Signal Processing
• Machine Learning
• Ocean Acoustics

Background:

• Conventional sonar beamforming methods like delay-and-sum exhibit limitations in resolution and performance under low signal-to-noise ratio (SNR) and interference.
• Existing beam-domain deconvolution techniques struggle with ill-posed inversions, noise sensitivity, and shift-variant point-spread functions (PSFs) in wide-field imaging.

Purpose of the Study:

• To develop an advanced deconvolution beamforming method that overcomes the limitations of conventional techniques.
• To enhance the resolution and robustness of sonar imaging in challenging underwater acoustic environments.

Main Methods:

• A noise-aware deconvolution beamforming convolutional neural network (NA-DBF-CNN) was proposed, integrating a fully convolutional encoder-decoder architecture.
• The network incorporates a physics-guided consistency constraint and is trained with a hybrid objective, including peak-emphasized supervision and SNR-weighted beam-domain data-consistency loss.
• Noise-aware weighting accounts for sample-dependent reliability in mixed-SNR conditions.

Main Results:

• The NA-DBF-CNN demonstrated improved robustness and alleviated reliance on shift-invariant PSF approximations.
• Simulations and lake experiments confirmed enhanced performance in multi-target scenarios, particularly under low SNR and coherent interference.
• The method effectively addressed the ill-posed nature and noise sensitivity of traditional deconvolution approaches.

Conclusions:

• The proposed NA-DBF-CNN offers a significant advancement in sonar imaging by improving resolution and performance.
• This physics-guided deep learning approach provides a more robust solution for underwater acoustic imaging, especially in complex and noisy conditions.