Underwater object detection is one of the most challenging research topics in computer vision. The complex underwater environment introduces issues such as high noise levels, low visibility, blurred edges, low contrast, and color distortion, which present significant challenges for underwater object detection tasks. To address these challenges, this study proposes an improved method based on the YOLOv8n model. Firstly, to tackle the problem of detecting blurred small objects, we propose the Blurred Small Object Enhance Pyramid (BSOEP), an improvement on the original PAFPN, which effectively captures multi-scale feature representations from global to local levels, allowing the network to better detect blurred small objects. Secondly, we integrate the ADown-Conv module into the network to reduce model complexity by lowering the number of parameters. This design maintains as much image information as possible while reducing the spatial resolution of feature maps, thus enhancing model efficiency without compromising detection accuracy. Thirdly, we replace the CIoU loss function with GIoU to improve robustness against changes in object shape and scale, enabling the model to better focus on the contributions of overlapping and non-overlapping parts of bounding boxes, thereby adapting more effectively to variations in object size and position. We evaluate the proposed algorithm on the URPC and DeepFish datasets. Compared to YOLOv8n, the mAP@0.5 on the URPC dataset increased by 0.6%, and the mAP@0.5-0.95 increased by 1.5%. On the DeepFish dataset, the mAP@0.5 increased by 0.6%, and the mAP@0.5-0.95 increased by 1.4%. Based on the URPC and DeepFish datasets, the experiments demonstrate that our improved model outperforms the current state-of-the-art models for underwater object detection.