Lightweight sonar target detection based on dual-domain hierarchical evolution

Existing sonar image target detection methods typically extract high-dimensional features within the spatial domain while ignoring frequency domain information. This leads to the coupling of detail and contour features, which in turn causes blurred target edges and reduces detection accuracy. Meanwhile, existing network structures are complex and struggle to balance the lightweight and real-time requirements of resource-constrained underwater platforms. To address these issues, this paper proposes a lightweight sonar target detection method based on dual-domain hierarchical evolution. The method divides the backbone network into multiple layers, each handling distinct extraction tasks, namely high-resolution detail preservation, frequency domain enhancement, and high-level semantic learning. Through spatial-frequency dual-path collaborative modeling, the method enhances target structure and edge representation while maintaining a lightweight architecture. Experiments on three public sonar datasets MDD, WHD, and UATD show that the proposed method improves mAP 50 : 95 by 1.3%, 1.0%, and 1.6% over the second-best methods, respectively. It has only 1.2 M parameters and achieves 122.1 FPS, outperforming state-of-the-art models. This study validates the effectiveness of dual-domain hierarchical evolution in sonar target detection and offers an efficient solution for real-time detection on resource-constrained underwater platforms. • A dual-domain hierarchical evolution paradigm for lightweight sonar target detection. • A spatial-frequency dual-path (SFD) module with frequency attention and multi-scale receptive fields. • HierNet achieves 41.2% mAP 50 : 95 on UATD with only 1.2 M parameters. • Inference speed reaches 122.1 FPS, enabling real-time deployment on resource-constrained platforms.