Abstract: Due to the limited underwater sonar image data and supervisory information for underwater targets,existing object detection algorithms are challenging to apply directly.To address this issue,this paper proposes an open-set underwater sonar image object detection method,USD(Underwater Sonar Detection),based on DETR.In the cross-modal feature fusion encoding module,it employs a multi-scale deformable attention mechanism to process image features iteratively,enabling the network to selectively focus on important information while reducing computational load.Simultaneously,it designs a multi-head self-attention mechanism to iterate text features,enhancing the model’s global modeling capability for sequences.Next,it utilizes a bidirectional attention mechanism to fuse text and image features,emphasizing the bidirectional relationships within the input sequences and enabling the network to capture more complex text-image interactions.Additionally,in the image-text feature decoding module,it uses image features to initialize queries,which are output from the Encoder module,and applies the DN method to address the issue of slow model convergence during training.Experiments show that the proposed method achieves a mean average precision of 77.5% on a custom underwater sonar image dataset,outperforming other detection methods in terms of precision,meanwhile successfully implements open-set object detection with robust performance.

Key words: Deep learning, Underwater sonar images, Open-set object detection, Feature fusion, Multimodal