关键词: 红外舰船检测, YOLOv8, GSConv, BiFPN, 图像处理

Abstract: To address the challenges of low image contrast, weak target features, missed detections of small distant targets, and the redundancy of existing detection models in infrared ship detection under complex nighttime sea conditions, this paper proposes a multi-scale lightweight object detection network—GRBP-YOLO. First, the Group Shuffle Convolution (GSConv) module is introduced to replace the traditional convolutional layers. By leveraging channel grouping and feature rearrangement mechanisms, GSConv achieves a balance between computational efficiency and representational capability, significantly reducing model complexity while maintaining detection accuracy. Second, a Residual Split Cross-Stage Partial Fusion Module (RS_C2f) is proposed. Through the synergistic design of residual connections and a split-stacking structure, RS_C2f strengthens cross-layer feature interaction while inheriting the lightweight advantages of the original C2f module, further reducing parameter redundancy. Finally, a Bidirectional Feature Pyramid Network (BiFPN) is constructed, incorporating an additional P2-level small object detection head. This design enhances the detection of tiny targets by recalibrating multi-scale features, effectively mitigating the missed detections of small, low-contrast ship targets at long distances caused by limited resolution. Experimental results show that, compared to the baseline model YOLOv8n, the GRBP-YOLO algorithm achieves an average improvement of 1.5% and 1.9% in detection accuracy (mAP@0.5) and recall rate, respectively, reaching 92.5% and 88%. The model parameter count is reduced to 1.7 million, and the file size is only 3.9 MB. The experiments validated the effectiveness of the GRBP-YOLO algorithm, providing a new technical solution for real-time ship detection in complex maritime conditions.

Key words: Infrared ship detection, YOLOv8, GSConv, BiFPN, Image proceessing