Underwater Object Detection Based on Improved Transformer and Attentional Supervised Fusion

Underwater object detection is one of the important technologies for improving efficiency underwater inspection, but existing methods still suffer from problems missed and insufficient target localization capability targets. To address these problems, an improved Transformer multi-scale attentional supervised feature fusion-based method proposed. In our method, objects are preprocessed by prior knowledge first. Then, a new coordinate decomposition window-based (CDW) block proposed to extract spatial location information more accurately, scaling factors introduced reduce intermediate computation. Finally, fusion (ASF) strengthen link between extraction fusion, further improve detected performance using compound attention weights. The cascade head improved, where flow reversed enhance prediction coordinates. average accuracy on URPC DUO datasets 3.7% 3.8% higher than that baseline network through cross-test, outperforms state-of-the-art methods. This study can provide reference engineering applications such as automated marine operations biodetected fishing techniques.