Attention-Enhanced U-Net for Finding Drowned Victims Underwater Using Sonar Equipment

Time and accuracy are key elements of underwater search and rescue operations. This article proposes a system with a modified U-net architecture, leveraging spatial and channel attention techniques for binary segmentation of drowned victims from side-scan sonar imagery. The base architecture is enhanced by a novel spatial skip-connection attention (SSCA) and multihead attention (MHA). The SSCA improves feature representation by highlighting relevant spatial information in the shallow layers. At the same time, the MHA module captures abstract channel-wise relations, improving the information flow and feature representation deeper in the network. In our experiments, we used original samples collected from the Oder River in Szczecin, using a drowned dummy and an Edgetech 4125 for sonar imaging, to produce a novel sonar drowned victims data set. The results obtained on the test set (99.54% accuracy and 75.76% mean dice coefficient) show that the model can adapt to the task, achieving significant results.