Automatic sonar target recognition using regularized wavelet neural networks

Convolutional Neural Networks (CNNs) have gained prominence in the sonar signal processing community due to their ability to capture complex patterns in acoustic data. However, in automatic sonar target recognition, the limited data availability often leads to overfitting, making network training challenging. Recent studies have shown that CNNs tend to learn filters that resemble wavelets, suggesting that incorporating wavelet-based priors into the design of convolutional neural networks could reduce the learning burden. Building on this insight, we propose a wavelet neural network that replaces learned filters in a CNN with a tight wavelet frame, learning only how to combine wavelet scales. Combining a tight wavelet frame with regularization constraints provides effective control over overfitting of the proposed wavelet neural network. The proposed approach is evaluated on real-world sonar target echo return datasets. [This talk will present research funded by DoD Navy (NEEC) (Grant No. N00174201001) and the ONR (Grant numbers N000142112420 and N000142312503)]