Sonar Object Detection Based on Global Context Feature Fusion and Extraction

Automatic underwater target detection is a key task of sonar image processing and analysis, which aims to identify the target category and achieve accurate localization. However, sonar images are affected by seabed reverberation noise and complex background, resulting in great challenges for sonar target detection. YOLO network has shown excellent performance in the field of object detection, but its network structure is too complex, and its detection effect on sonar targets is not good. To address these difficulties, this paper proposes a sonar image target detection network based on Global Context Feature Fusion and Extraction (GCFFE-Net), which can globally extract and fuse semantic features from sonar images at multiple scales, and bidirectionally transfer features between different levels, thus highlighting the target area. Specifically, the multi-scale channel feature fusion module (MCFF-Block) in the network uses dilated convolution, and enhances the semantic information interaction between the feature layers required for different anchor detection by adaptively adjusting the weights of different channels. The spatial context feature pyramid fusion module (SCPFF-Block) adopts a pyramid structure, and realizes the fusion of shallow features and deep features of different levels of size through bidirectional information transfer from top to bottom and from bottom to top, improving the network's perception ability of the spatial context information of the target. The hierarchical region of interest attention module (HAA-Block) aligns and pools the regions of interest of different levels of size, making the network more focused on the location of the target, and enhancing the accuracy and robustness of target detection. The paper conducts experiments on the public dataset UADT, and the results show that the network proposed in this paper has high performance and superiority in sonar image target detection.