Acoustic Simulated Sonar Data and Neural Networks for Marine Mine Detection

Underwater object detection has always been a particular challenge due to the unique physical context of the underwater environment. Nevertheless, this challenge remains essential for states concerned with protecting coastal areas from the threat of underwater mines and monitoring the seabed. The notion of“object detection” refers to the process of detecting and classifying an object present in an image using a reference box. Object detection from a sonar image is now largely carried out using algorithms based on deep neural networks. Indeed, given the vast areas to cover and the limited number of operators, the need for automation, or at least semi-automation of the process, is inevitable. The training of these networks requires large and diversified annotated database to ensure effective learning and a reliable behaviour. However, rapidly constituting such a database is challenging due to the high time and resource cost of multiplying acquisition campaigns. In this context, the simulation of sonar images is an interesting option to compensate for this lack of data. This paper presents a new protocol for SAS images simulations to train object detection models. This project stands out for two main aspects: on one hand, the 3D modelling of the seabed from underwater photogrammetry acquisitions, and on the other hand, the optimization of the parameters of a sonar image simulator to increase the realism of the images. This massive production of simulated data will then serve to feed an object detection model through the implementation of domain adaptation methods.