Aim

This project aims to develop techniques for extracting information from sonar data, leveraging physics-aware neural network architectures to overcome biases and other sources of errors in the raw data.

Objectives

1. Model sonar signal reconstruction as an inverse problem based on physics. Special attention will be given to the modelling of noise and biases in the signal acquisition.
2. Explore and develop techniques based on physics-aware neural networks and differentiable renderers to solve the sonar signal reconstruction inverse problem.
3. Benchmark the developed techniques on relevant real and synthetic data.

Description

Sound navigation and ranging (sonar) is used in a wide range of underwater applications, from archaeology and environmental monitoring to pipe inspection and defence. In active sonar, a transmitter emits sound waves that are reflected by objects of interest, the seabed floor, and other elements, and used by the receiver to reconstruct an image of the environment, which is typically 2D, but its dimensions can vary depending on the sensor array. Being underwater, however, introduces several challenges, including noise and multipath reflections, making the processing of sonar data challenging. Reconstructing a 3D scene from 2D sonar data, in particular, is a difficult ill-posed inverse problem [1-2].

In this project, we draw on recent progress in physics-aware neural networks [3-4] and differentiable physics/graphics rendering [5] to tackle this problem. Specifically, to improve generalisation, robustness, and data efficiency, we devise strategies to embed the physics of sonar into neural networks that solve the reconstruction problem and parameterise the object of interest with a differentiable physics renderer. The resulting algorithms will be illustrated on relevant synthetic and real data. In particular, synthetic data will be generated realistically, considering complexities such as environmental effects (noise and multipath) and sensor noise and inaccuracy (resolution, biases and errors).

Further information

References

[1] M. D. Aykin, S. Negahdaripour, "Forward-look 2-D sonar image formation and 3-D reconstruction", IEEE OCEANS, pp. 1-10, 2013.

[2] T. Guerneve, K. Subr, and Y. Petillot, "Three‐dimensional reconstruction of underwater objects using wide‐aperture imaging SONAR", Journal of Field Robotics, vol. 35, no. 6, pp. 890–905, Sep. 2018, doi: 10.1002/rob.21783.

[3] J. Willard, X. Jia, S. Xu, M. Steinbach, and V. Kumar, "Integrating Scientific Knowledge with Machine Learning for Engineering and Environmental Systems", ACM Comput. Surv., vol. 55, no. 4, pp. 1–37, Apr. 2023, doi: 10.1145/3514228.

[4] C. Banerjee, K. Nguyen, C. Fookes, and K. George, "Physics-Informed Computer Vision: A Review and Perspectives", ACM Comput. Surv., vol. 57, no. 1, pp. 1–38, Jan. 2025, doi: 10.1145/3689037.

[5] A. Spielberg et al., "Differentiable visual computing for inverse problems and machine learning", Nat Mach Intell, vol. 5, no. 11, pp. 1189–1199, Nov. 2023, doi: 10.1038/s42256-023-00743-0.

Access to data:

Through working with the Dstl partner, there will be opportunities to gather data under multinational trials opportunities with the NATO Centre for Maritime Research and Experimentation (CMRE). A set of trials entitled “Cognitive Sonar” are anticipated to take place from 2026, and Dstl are able to influence the trials plan and potentially include serials with specific data collection that assists with the PhD project goals, provided there is general alignment with the multinational objectives. Some existing datasets may also be available and sharable with the project. As part of this work, there may be an opportunity for the PhD student to participate in trials planning meetings and even the trial itself. Throughout the project and particularly in the lead up to trials planning, it will be beneficial for the student to consider what data collection could be useful and suggest trial serials that achieve