NFFLS: Rapid and Accurate Underwater 3-D Reconstruction With Neural Fields for Forward-Looking Sonar

Forward-looking sonar (FLS) can capture high-resolution acoustical images from the underwater scenes, maintaining performance even in turbid water conditions and poor lighting. Although neural fields have become popular for 3-D reconstruction from acoustical images, they still suffer from poor accuracy and slow computation. We thus propose a novel framework, namely, neural fields for FLS, which leverages a combination of advanced encoding strategies and tailored loss functions for rapid and accurate object-level 3-D reconstruction. The framework introduces multiresolution hash encoding combined with frequency positional encoding to efficiently represent spatial features, thereby enhancing the capture of scene characteristics. A novel ray sampling strategy is also developed to improve training efficiency while preserving the reconstruction quality. In addition, the proposed framework incorporates a target encoding approach alongside a classification loss to robustly handle noise-rich sonar data, while a structural intensity constraint ensures consistent capture of the intrinsic features of sonar images. Quantitative and qualitative validations using simulated and real scenarios demonstrate substantial improvements in both reconstruction accuracy and training efficiency over existing state-of-the-art methods. The combination of these strategies forms a unified framework that addresses the challenges, providing a more stable, detailed, and computationally efficient solution for FLS-based 3-D reconstruction.