DSM-Net: A multi-scale detection network of sonar images for deep-sea mining vehicle
Xinran Liu, Jianmin Yang, Wenhao Xu, Qihang Chen, Haining Lu, Yu Chai, Changyu Lu, Yulong Xue
Abstract
• This study introduces DSM-Net, a deep learning-based sonar image detection network designed for deep-sea mining vehicles, addressing challenges like multi-scale detection of seabed terrains and noise interference. • The network introduces two innovative modules, TSA and PDM, which enhance multi-scale feature extraction, reduce noise, and improve inference speed and detection accuracy. The proposed ADL function effectively addresses the issue of target imbalance. • Extensive experiments and sea trials validate the superior performance of DSM-Net in sonar image detection, ensuring operational and navigational safety for deep-sea mining vehicles, and supporting subsequent path planning and obstacle avoidance tasks. Deep-sea mining vehicles (DSMVs) play a crucial role in deep-sea mining operations, requiring high-precision, real-time detection of seabed rocks and terrain of varying scales to ensure safe navigation and operation. However, the complexity of multi-scale seabed terrains, along with the low resolution and high noise levels in sonar images, makes accurate real-time detection a challenge. To address these issues, DSM-Net, a multi-scale terrain detection network specifically designed for deep-sea mining, is proposed. DSM-Net integrates several innovative modules: the Tri-Scale Attention Module (TSA) extracts multi-scale features and reduces noise interference, the Partial-Dynamic Module (PDM) improves inference speed, and the ASFF* detection head incorporates an additional small-target detection layer. Furthermore, an adaptive weighting function, Adaptive Difficulty Loss (ADL), is introduced to handle the imbalance in the number of targets across different scales in actual seabed environments. DSM-Net was evaluated on the DSMSD dataset, showing a 2.16 % improvement in [email protected]:0.95 and an 18.75 % reduction in inference time compared to the YOLOv8 baseline, striking an effective balance between detection speed and accuracy. In sea trials, DSM-Net contributed to path planning and obstacle avoidance, proving its practical engineering value.