SonarNet: A Real-time Anchor-free Detector for Sonar Object Detection

Due to the weak illumination in deep water areas and the complex underwater environment, captured images often contain a variety of irregularly shaped objects. This makes it difficult for traditional visual inspection methods to achieve accurate detection of underwater objects. Therefore, we propose a real-time anchor-free sonar object detection network, SonarNet. First, to extract fine local features in high-resolution image layers, we propose the attention pyramid pooling module (APP), which amplifies and enhances the features of localized objects in the low-order features. Second, the multi-head attention convolution module (MAC) is introduced to improve the network's overall understanding of the information. Finally, the semantic refinement fusion module (SRF) is implemented to enhance the network's grasp of global information and reduce noise interference from the complex background. We evaluate the performance of our proposed SonarNet using the public dataset UATD and compared it with other state-of-the-art methods. Notably, SonarNet achieves a mAP of 0.878, outperforming all other methods and delivering superior performance in most categories.