Dual-branch framework: AUV-based target recognition method for marine survey

Abstract

Autonomous recognition of marine targets is considered a promising technology for autonomous underwater vehicle (AUV) marine survey, and AUV equipped with side-scan sonar (SSS) for recognition is the key to surveys. As a fundamental function, SSS recognition remains unsolved due to the challenging image conditions of SSS and insufficient algorithm robustness. This paper proposes an accurate and real-time dual-branch recognition framework containing segmentation and refinement branches. Firstly, the segmentation branch uses a lightweight learning network to analyze the data comprehensively. In this branch, we propose a densely connected local attention recurrent residual (LAR2) block as the backbone, and at the same time, an atrous convolution is introduced. This branch can focus on the features of interest in the image, ensuring better feature representation with low-resolution SSS information while guiding the next branch. Secondly, the refinement branch is to adjust the previous branch’s results and combines the low-level and high-level features. We propose holistic attention (HA) block in this branch, which can further improve the target recognition performance. Finally, we adopt the feature fusion method of bilinear pooling to integrate the results of the two branches to output a high-precision recognition image. In offline experiments and sea trials, our proposed method outperforms other competing algorithms in the four indicators of semantic segmentation, and achieves a computation speed of 92.66 ms (±0.86 ms) per image on AUV dedicated hardware. The method has strong robustness, meets real-time performance, and can be widely used in AUV marine survey.