A Deep Learning Method for UUV to Detect Undersea Landform Features

Marine topography refers to the configuration and composition of underwater landscapes, which holds significant importance for marine resource exploration and environmental research. However, conventional methods for detecting terrain and geomorphology in marine environments often struggle to accurately identify these features due to the intricate nature of the marine setting and data interference. To tackle this challenge, a proposal is made to leverage deep learning techniques for exploring ocean terrain and geomorphology. As the utilization of Side-scan Sonar (SSS) in underwater settings becomes increasingly prevalent, the study of sonar image processing is becoming more comprehensive. This study introduces a feature extraction approach for Unmanned Underwater Vehicle (UUV) SSS images based on a multi-layer feature fusion structure of a lightweight Fully Convolutional Network (FCN). The method aims to address issues such as low precision in feature extraction, inadequate boundary coherence, and suboptimal edge detail in existing methods for extracting features from SSS images. Recognizing the abundance of convolution parameters in the fully connected layer leading to network model complexity and information loss, a lightweight FCN is devised to reduce computational load, simplify the network model, and prevent information loss due to large dimension spans. Experimental findings demonstrate that this method not only ensures adequate network training but also leverages information correlation across different convolution layers to enhance the network’s ability to capture and fuse effective feature information, thereby improving the integrity of feature extraction edges in SSS images.