Underwater target detection algorithm fusion with multiple attention mechanisms and Wise-IoUv3

To address the challenges of image blurring and low detection accuracy in complex underwater environments, this study proposes a novel underwater target detection algorithm that combines multiple attention mechanisms with the Wise-IoUv3 loss function. Initially, a multi-scale feature enhancement mechanism is developed, where Omni-Dimensional Dynamic Convolution (ODConv) is employed to replace certain convolutional layers in the backbone network. Additionally, an efficient multi-scale attention (EMA) mechanism is integrated to enhance the network's ability to extract features from blurred and small targets.Furthermore, the Spatial Pyramid Pooling (SPPF) module is optimized by incorporating an average pooling branch, which enriches spatial information and strengthens global context awareness. To further improve efficiency, a lightweight BiFormer attention mechanism is integrated into the dual-branch structure, reducing computational complexity while enhancing the detection performance for small targets.During the prediction phase, the original loss function is replaced with Wise-IoUv3 to better balance the training outcomes for images of varying quality. Finally, the conventional detection head is substituted with a dynamic detection head (DynamicHead), which enhances scale perception, spatial perception, and task perception capabilities, thereby improving the accuracy of object localization.Experimental evaluations on the RUOD and URPC datasets demonstrate that the proposed algorithm achieves a 3.6% and 1.7% improvement in average precision compared to the baseline method. The results indicate that the proposed approach significantly reduces missed detections and false positives, leading to enhanced overall detection performance.