Recognition of Underwater Engineering Structures Using CNN Models and Data Expansion on Side-Scan Sonar Images

Abstract

Listen

Translate article

Side-scan sonar (SSS) is a critical tool in marine geophysical exploration, enabling the detection of seabed structures and geological phenomena. However, the manual interpretation of SSS images is time-consuming and relies heavily on expertise, limiting its efficiency and scalability. This study addresses these challenges by employing deep learning techniques for the automatic recognition of SSS images and introducing Marine-PULSE, a specialized dataset focusing on underwater engineering structures. The dataset refines previous classifications by distinguishing four categories of objects: pipeline or cable, underwater residual mound, seabed surface, and engineering platform. A convolutional neural network (CNN) model based on GoogleNet architecture, combined with transfer learning, was applied to assess classification accuracy and the impact of data expansion. The results demonstrate a test accuracy exceeding 92%, with data expansion improving small-sample model performance by over 7%. Notably, mutual influence effects were observed between categories, with similar features enhancing classification accuracy and distinct features causing inhibitory effects. These findings highlight the importance of balanced datasets and effective data expansion strategies in overcoming data scarcity. This work establishes a robust framework for SSS image recognition, advancing applications in marine geophysical exploration and underwater object detection.