Implicit Frequency-domain Acoustic Fields for Underwater Reconstruction

Underwater reconstruction algorithms are commonly categorized into two types: reliant on optical data and reliant on acoustic data. The propagation of light underwater is subject to Beer-Lambert’s law, resulting in significant attenuation, particularly in realistic complex environment. Consequently, in the reconstruction of vast and intricate underwater environments, acoustic-based methods predominate. The effectiveness of acoustic imaging and positioning is directly linked to both the quality and quantity of acoustic signals obtained. However, acquiring extensive acoustic signals for underwater reconstruction can be financially prohibitive. Hence, there arises a critical need to develop methods for learning implicit representations of the acoustic field, facilitating data augmentation, and enhancing the efficiency of underwater reconstruction processes. In this work, inspired by neural radiance fields (NeRFs), we focus on harnessing the potential of the mel spectrogram as a dataset for learning implicit frequency-domain acoustic fields, named Mel Spectrogram Implicit Acoustic Field (MSIAF). By encoding large-scale time-domain information into small-scale frequency-domain information, MSIAF is able to quickly learn an implicit representation of the sound field in the target space, and then obtain the response function at any location. We use MSIAF to optimise the elliptical inverse projection underwater reconstruction method to demonstrate the possibilities of MSIAF for underwater reconstruction algorithm optimisation. Our experiments reveal that MSIAF exhibits substantial improvements over Neural Acoustic Fields(NAF) in specific aspects, underscoring its efficiency and potential for enhancing acoustic field representation.