SExtract-Net: A Lightweight Multi-Scale Feature Fusion Network for Sonar Image Recognition

Sonar imaging plays a vital role in underwater sensing, with broad applications in ocean exploration and intelligent object recognition. However, sonar images often suffer from low resolution, blurred edges, and complex background noise, which significantly hinder recognition accuracy and robustness. Most existing models rely on general-purpose architectures that fail to fully exploit sonar-specific features, making it difficult to extract discriminative information from low-contrast, multi-scale targets. To tackle these challenges, this paper proposes SExtract-Net, an efficient sonar image recognition model based on PPHGNetV2. It first introduces the Stem-ASPP module, which incorporates Atrous Spatial Pyramid Pooling (ASPP) to capture multi-scale contextual features and enhance both local and global perception. The paper then develops a lightweight feature extraction module, SwiftExtractNet (SExtract), and constructs an enhanced block, HGV2-SExtract, to improve multi-level feature fusion and representation. Extensive experiments on a combined dataset from multiple public sonar sources demonstrate the model's effectiveness.