Underwater Object Detection Algorithm Based on Improved YOLOv12

Sonar acoustics, with their low transmission loss and immunity to factors like optical visibility, have become the primary means for underwater target detection. However, traditional sonar image recognition technologies face numerous challenges. Therefore, this paper proposes an underwater target detection algorithm named TGC-YOLOv12. First, a GhostConv module that combines attention mechanisms with convolution is added to achieve a lightweighting model while maintaining recognition accuracy. Next, the TripletAttention module is incorporated into the backbone network to enhance useful features and suppress irrelevant ones through feature recalibration, which helps the detection model recognize and locate targets more accurately. Then, in the neck part of the network, the CCFM cross-scale fusion is modified to improve model performance while maintaining lightweight characteristics. Experimental results on our dataset show that TGC-YOLOv12 achieves a mean Average Precision ([email protected]) of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3 1. 8 \%}$</tex>, with model parameters totaling <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$9.8 \times 10^{6}$</tex>. Compared with other state-of-the-art underwater object detection algorithms, TGCYOLOv12 not only achieves higher detection accuracy but also reduces computational overhead.