Multiscale part-based feature fusion network for fine-grained ship detection in remote sensing images

Fine-grained ship detection in optical remote sensing imagery serves as a fundamental technology for maritime traffic control and fishery management. This task encounters three primary difficulties: the presence of geometrically complex ship structures, significant visual similarity among different ship categories, and repeated structural patterns across vessel types. To overcome these challenges, we present a multipart feature fusion framework comprising three core components: the midpoint offset strategy (MOS) module, the part feature fusion module, and the multiscale part-based feature relation (MPFR) module. To address the computational complexity and inaccuracy of traditional angle regression methods, we introduce an MOS module that replaces angle parameters with six length-based measurements, transforming angle regression into length regression for more accurate rotated proposals. A part feature fusion module is then proposed, incorporating saliency loss and exclusivity loss to enhance feature representation. The saliency loss guides the network to focus on the most discriminative regions, whereas the exclusivity loss encourages different channels to attend to distinct salient areas, improving information extraction. Finally, an MPFR module is developed, which utilizes multiscale sliding windows to extract local details and enhance their dimensional representation. These features are refined through transformer blocks to model inter-scale relationships, improving classification performance. Experiments on the HRSC2016 dataset demonstrate the effectiveness of the proposed method for fine-grained ship detection in complex maritime environments. These refined feature representations are subsequently fed into the final detection head to perform target regression and classification. Experimental validation on the HRSC2016 benchmark confirms our framework’s effectiveness, attaining 86.82% mean average precision. This performance exceeds current rotation-based detection methods by 3 to 4.2 percentage points, demonstrating substantial improvements in fine-grained ship recognition.