Deep Learning-Based Fourier Registration for Forward-Looking Sonar Odometry in Texture-Sparse Underwater Environments

Abstract

Robust forward-looking sonar (FLS) odometry is critical for underwater autonomous navigation but is hindered by severe noise and sparse textures in acoustic imaging. Traditional Fourier-based methods are susceptible to such degradations, while end-to-end deep learning approaches often struggle to learn intrinsic geometric relationships. We propose a novel deep learning framework that synergizes classical signal processing with learnable architectures. Our method decomposes pose estimation into rotation and translation stages, utilizing an improved Trans-UNet to enhance image feature interaction. Specifically, the rotation network leverages the Radon transform for noise filtering, combined with a multi-angle correlation layer to determine angular relationships. Following rotation correction, an improved learnable phase correlation module estimates translation within an end-to-end trainable system. Experiments on public datasets demonstrate that our method achieves outstanding odometry performance even without loop closure detection, and zero-shot evaluations on wetland datasets further validate its strong generalization capability.