SDNet: Image‐based sonar detection network for multi‐scale objects

Abstract

Autonomous Underwater Vehicle (AUV) carrying sonar for object detection has become one of the main ways of ocean exploration. However, object detection in sonar images always faces the problems of balancing detection accuracy and efficiency. To this end, this paper proposes an efficient underwater object detection network for sonar images named SDNet. In the model, the authors construct a new feature extraction network based on RepVGG to balance the detection accuracy and speed. By combining channel attention with RepVGG, useful information of high-order feature maps captured can be selectively paid attention to. Then, the authors design a feature fusion network to efficiently converge the location and semantic information of multi-scale feature maps. In the network, the authors propose a lightweight cross stage partial network for sonar (CSP_S) module suitable for sonar images, which can enhance the model's feature fusion capability and simplify the model. Finally, to reduce the conflict between classification and regression tasks, the authors leverage the Decoupled Head for the sonar object classification and localization. By testing on the self-built Underwater Sonar Dataset underwater sonar dataset (USD) and the public sonar dataset sonar common target detection dataset (SCTD), the detection accuracy of SDNet reaches 99.52% and 95.20%, respectively. Moreover, the detection speed reaches 114 frames per second (FPS) and 138 FPS, respectively. The experimental results show that SDNet can effectively balance the sonar detection accuracy and efficiency.

CCDN-DETR: A Detection Transformer Based on Constrained Contrast Denoising for Multi-Class Synthetic Aperture Radar Object Detection

“…Researchers have attempted to adapt and apply these methods to SAR target detection by making specific improvements. These improvements mainly include using stronger backbone networks [ 8 , 9 ], setting up multi-scale FPN layers [ 10 , 11 ], and designing loss functions more suitable for SAR tasks [ 12 , 13 ]. Meanwhile, real-time SAR target detection schemes [ 14 , 15 , 16 ] are also gradually developing, providing references for the practical application of SAR detection and recognition.…”