Underwater Targets Recognition Based on Multiple AUVs Cooperative via Recurrent Transfer-Adaptive Learning (RTAL)

Abstract

The recognition of multiple Autonomous Underwater Vehicles (AUVs) equipped with Side-Scan Sonar (SSS) is the key to marine surveys, and autonomously and efficiently recognizing marine targets is an urgent open challenge. This paper proposes a lightweight Recurrent Transfer-Adaptive Learning (RTAL) to improve the recognition accuracy of SSS images. In this work, (1) we analyze the target features collected by multiple AUVs, calculate the similarity between features using Mahalanobis distance, and establish a threshold. (2) When the threshold is higher than the similarity, AUVs adopt improved Recurrent Transfer Learning (RTL) to reduce computational consumption and ensure the real-time performance of the algorithm. (3) When the similarity is lower than the threshold, AUVs adopt RTAL to reconstruct the data of the target information with inconspicuous features caused by the complex background, extract features efficiently, and reduce the interference of environmental factors. We compare several classical neural networks and validate them on dataset collected from AUVs sea trials. The experimental results show that our proposed method has high recognition speed and recognition accuracy.