Underwater acoustic image target detection based on bidirectional fusion pyramid and channel attention

For real-time and accurate detection of underwater objects, this paper proposes an underwater acoustic image target detection model, the bidirectional pyramid channel attention network (BPCA-Net), which incorporates a bidirectional pyramid network and a channel attention network to achieve end-to-end underwater object classification and localisation. First, the BPCA-Net model incorporates a channel attention mechanism into the advanced backbone network ConvNeXt to recombine the extracted feature map channels to highlight useful object feature information. Then, the feature maps are fed into a designed lightweight bidirectional fusion pyramid network for multi-scale feature fusion and enhancement. Finally, the multiple feature maps are fed into a single stage detection head for classification and localisation. The experimental results show that the proposed model can extract important features of objects in underwater acoustic images, demonstrating high accuracy compared to existing detection algorithms.