To solve the problem of occlusion and false and missed detection caused by insufficient illumination and noise interference in the underwater environment, as well as the aggregation of small targets, this study proposes an improved underwater target detection model based on YOLO11n. Meanwhile, the wavelet pooling convolutional network (WPCN) is put forward to decompose features by wavelet pooling technology and preserve low-frequency subbands, thus avoiding frequency aliasing. Pruning optimization is adopted to improve detection accuracy while maintaining inference speed. Additionally, single head visual Transformer (SHSA) and convolutionally gated linear module (CGCM) are introduced to improve the adaptability to complex scenarios. A shared reparameterizable convolutional detection head (RLD-Head) is proposed to reduce the amount of parameters by sharing the convolutional layers, and reparameterization techniques are utilized to avoid performance loss and adapt to resource-constrained devices. The Wise-inner-MPDIoU loss function is designed to improve the detection accuracy. Compared with the original model, mAP50 improves by 3.8 percentage points, mAP50-95 increases by 4.3 percentage points, and the parameter amount and computation amount are reduced by 30.6% and 30.1% respectively, which proves the advantages of this innovation in underwater target detection.