An Improved YOLOv11-Based Recognition Method for Side-Scan Sonar Images of Shipwrecks

With the rapid advancement of underwater autonomous vehicle (AUV) technology, AUVs equipped with sidescan sonar is capable of autonomously recognizing targets within sonar image and conducting re-exploration operations based on the recognition outcomes. However, sonar images of underwater targets primarily reflect the echoes intensity and distance of the objects, resulting in sparse feature representations. Additionally, the dataset for target images is relatively limited. Direct application of traditional target recognition algorithms would likely yield suboptimal performance due to these constraints. To address this issue, after the backbone of YOLOv11n, the deformable bi-level attention module is used to enhance the model's ability to focus on critical features within the sonar images, thereby improving the accuracy of target recognition. We evaluated the proposed model using the open-source SeabedObjects KLSG dataset, achieving an mAP@ 0.5 of 96.53% and an [email protected] of 53.13%. These results represent improvements of 2.3 % and 8.36 %, respectively, over YOLOv11n.