Lightweight sonar model targets small underwater objects at long range

Researchers have proposed a lightweight target detection model designed to spot small underwater objects at long distances, combining an enhanced YOLOv5 detector with 2D image sonar hardware that is small in size, low in power consumption and easy to install. The work, published in Neural Computing and Applications, addresses a persistent difficulty in underwater sensing: detecting small targets reliably when they are far from the sensor.

The model modifies the YOLOv5 backbone with a self-attention mechanism, which the authors say strengthens the capture of features from distant targets. To keep the system practical for embedded deployment, the team applied model compression, enabling fast detection and localization of underwater small targets without heavy computational demands.

Beyond 2D detection, the approach fuses sonar data with target acoustic features to perform 3D reconstruction, allowing fine-grained detection of small targets.

Experimental results reported by the authors indicate that the proposed model accurately realizes long-distance detection of small underwater targets.

The combination of attention-enhanced feature extraction, model compression and acoustic 3D reconstruction is the paper’s central technical proposal.

Subject of Research: A lightweight sonar-based detection model for long-distance detection of small underwater targets.

Article Title: A lightweight target detection model for long-distance accurate detection of underwater small targets

Article References: Zhao, M., Zhao, Z., Shen, S., Li, R., & Ying, Z. (2026). A lightweight target detection model for long-distance accurate detection of underwater small targets. Neural Computing and Applications, 38(20), Article 837. https://doi.org/10.1007/s00521-026-12478-1

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12478-1

Keywords: underwater target detection, YOLOv5, self-attention, 2D image sonar, model compression, 3D reconstruction, acoustic features, sonar highlights, acoustic shadow, lightweight models, Neural Computing and Applications, lightweight