Underwater target detection is an essential technology in underwater operations. To address the challenges posed by complex backgrounds, varying target scales, and the presence of overlapping and occluded objects in underwater images, this study proposes an underwater target detection algorithm based on a dual-branch convolutional network. First, this study employs two parallel convolutional neural networks as the backbone, with one branch integrating the ECA attention mechanism and the other utilizing deformable convolutions to enhance the model’s feature extraction capability. Next, it utilizes the AFF module to effectively fuse the features extracted from both branches. Finally, the study adopts the PANet pyramid structure as the neck network to achieve multi-scale feature fusion while incorporating a high-resolution detection head to further improve sensitivity to small targets. Comparative experiments conducted on the publicly available underwater dataset RUOD show that the improved algorithm achieves an mAP50 of 86.8% on the RUOD dataset, which is an enhancement of 2.7% over the baseline YOLOv8n model. Moreover, this model outperforms other common target detection models of similar scale.