EDRNet: A Lightweight Underwater Sonar Target Detection Model Based on an Efficient De-Redundancy Network

Underwater object detection is one of the hot research topics in the field of acoustic signal processing. Using sonar to automatically detect seabed objects has significant importance for marine resource development and underwater military training. Traditional sonar image detection methods based on machine learning and clustering have problems such as large model parameters, redundant computation, low detection accuracy, and difficult deployment in practical engineering applications. To address these issues, this paper proposes a lightweight underwater sonar object detection model named EDRNet. First, it introduces the lightweight network EfficientFormerV2, which solves the problems of large model parameters and redundant computation. Second, it introduces the SimSPPF module, which enlarges the receptive field of the network. Finally, it adds the parameter-free attention mechanism SimAM, which enhances feature extraction and improves model computation efficiency and detection accuracy. Experiments on the SeabedObjects-KLSG dataset and the STCD dataset show that the proposed model achieves an average detection accuracy of 94.7%, which is 22.8%, 1.8%, 2.3%, and 0.4% higher than the existing mainstream models Faster R-CNN, YOLOv3, YOLOv4, and YOLOv5, respectively. The parameter size is 58.6% of YOLOv5s, and the computation amount is 51.25% of YOLOv5s. This verifies that the proposed model effectively reduces the model size while maintaining performance and has stronger practical significance for underwater sonar image detection.