Sea treasure target detection is a key technology for the intelligent development of sea treasure resources. This paper proposes an improved algorithm YOLOv9-PAEG based on YOLOv9-S to address the problem of low accuracy in detecting sea treasures in complex underwater environments, difficult feature extraction, diverse target sizes, and a large number of small targets. Firstly, the SPPELAN module was improved by introducing the PfAAM attention mechanism and distributed shift convolution DSConv2D, and the PFAD_SPPELAN module was designed to enhance the detection accuracy and speed of the model. Secondly, by introducing a variable kernel convolution AKConv in the backbone network layer of the model, the model can more flexibly adapt to features of different sizes and shapes, thereby improving its feature extraction ability for multi-scale targets, especially small targets. Then, the ECA attention mechanism was integrated into the neck layer of the model, enhancing its ability to represent important features and improving detection accuracy. Finally, by using the GIoU loss function, the convergence of the model was accelerated and the positioning accuracy was optimized. Experiments have shown that the YOLOv9-PAEG model performs well on datasets DUO and UDD mAP@0.5 They reached 89.7% and 77.6% respectively, and FPS reached 71 and 69, respectively. Compared with the original model and other mainstream object detection models, they have improved detection accuracy and speed. This fully proves the effectiveness and progressiveness of the YOLOv9-PAEG model, which can provide a better detection effect for marine treasures.