SSGA-YOLO: A Lightweight Sonar Image Object Detection Network With Efficient Convolution and Acoustic-Aware Attention for Embedded Systems

Abstract

To address the problems of high computational complexity and limited deploy ability in underwater sonar image detection, we propose SSGA-YOLO, a lightweight and efficient object detection algorithm optimised for sonar imagery and successfully deployed on the Ascend AI embedded platform. SSGA-YOLO focuses on three key aspects to achieve a favourable balance between accuracy and efficiency. The S-Net backbone employs depthwise separable convolutions and a lightweight attention mechanism to enhance the extraction of weak echo and shadow features in sonar images while reducing redundancy for efficient deployment. The Efficient Group Shuffle Convolution (EGSConv) enhances cross-channel feature interaction to improve the detection of small, low-contrast sonar targets and the Lightweight Shuffle-Aware Group Attention (LSGA) refines key acoustic and spatial cues in the presence of strong noise. Furthermore, SSGA-YOLO significantly reduces model parameters and computational complexity: compared to YOLOv8n, it achieves reductions of 79.82% and 74.07% in parameter count and GFLOPs, respectively. To evaluate model performance across diverse environments, embedded deployment experiments were conducted on three datasets representing distinct scenarios: MDFD for controlled artificial tanks, UATD for complex natural waters and MOTfish for dynamic video sequences. SSGA-YOLO consistently achieves high detection accuracy, with an mAP50 exceeding 0.930 on all datasets and peaking at 0.983 on MDFD. In terms of inference efficiency, the model demonstrates exceptional real-time capability, reaching a frame rate of 65.77 FPS on MOTfish. These results outperform other lightweight detectors, confirming the model's effectiveness for practical underwater applications.