Abstract: To address the insufficient fine-grained classification capability and suboptimal localization accuracy in current ship target recognition tasks, this study proposes an improved YOLO-based algorithm known as fine-grained classification and precise positioning YOLO (FP-YOLO). The algorithm introduces novel cross-stage partial bottleneck modules (D2f and G2f) into the backbone and neck networks, respectively, thereby enhancing both the detail feature extraction capability of the backbone and the global information interaction capacity of the neck. A backbone-neck feature fusion module (BNFM) was designed to effectively integrate same-level feature maps and optimize the complementary characteristics between the backbone and neck features, while improving the generalizability of the neural network. Additionally, a sliding loss function was introduced to better utilize the hard-to-classify samples during training. The experimental results demonstrated that the proposed algorithm achieved 66.9% mAP@0.95, outperforming YOLOv8n and YOLOv10n by 2.5 and 5.4 percentage points, respectively, on a custom-built horizontal fine-grained ship dataset. On a public remote sensing fine-grained ship dataset, the algorithm reached 57.2% mAP@0.5:0.95, an improvement of 3.4 and 7.0 percentage points, respectively, compared to the above two models.