DCB-YOLO: An Infrared Ship Detection Model with Deformable Convolution and Bi-level Routing Attention

Infrared detection technology has gained significant traction in maritime security systems, yet challenges persist in ship detection due to low-contrast and texture-deficient infrared images. To address the high complexity and inefficiency of existing models, this paper proposes an enhanced YOLOv8n-based algorithm for infrared ship detection. First, a Deformable Convolutional Network (DCNv2) is integrated into the backbone to adaptively capture multi-scale geometric deformations of ships. Second, a Bi-level Routing Attention mechanism (BiFormer) dynamically filters key regions, enhancing small-target detection in cluttered backgrounds. Third, a lightweight Bidirectional Feature Pyramid Network (BiFPN) optimizes cross-scale feature fusion while reducing redundancy. Finally, an Inner-IoU-based loss function refines bounding box regression by constraining intersection regions, improving localization for blurred edges. Experiments on the InfiRay dataset demonstrate that the proposed model achieves a mean Average Precision ([email protected]) of 90.1%, surpassing the baseline YOLOv8n by 2.3%. The model reduces parameters by 18.5%, computational load by 10.7%, and volume by 14.2%, achieving a compact size of 5.4MB. These advancements balance accuracy and efficiency, making the algorithm highly suitable for resource-constrained embedded deployment in real-time maritime surveillance. This work underscores the effectiveness of geometric feature modeling, dynamic attention routing, efficient multi-scale fusion, and precise localization strategies in overcoming challenges posed by complex marine environments, offering a robust solution for infrared-based ship detection.