RB-YOLO: enhanced multiscale object detection for routine maritime safety surveillance

Abstract

Visual object detection in complex nearshore environments faces challenges from wave clutter, illumination variation, and multi-scale objects, hindering maritime safety operations. This paper presents RB-YOLO, an enhanced framework based on YOLO11s. Key contributions: (1) We optimize the C3k2 module and integrate Receptive Field Attention Convolution (RFAConv) into the backbone to improve small and wave-distorted target detection by adaptively adjusting the receptive field. (2) We incorporate a Bidirectional Feature Pyramid Network (BiFPN) in the neck for robust multi-scale feature fusion under noisy visual conditions, and adopt the Mish activation function to enhance nonlinear expressiveness and gradient stability. (3) We employ Wise-IoU v3 (WIoUv3) loss to improve bounding box regression accuracy. Experiments on the AFO dataset show RB-YOLO achieves gains of 2.1% in precision, 1.2% in mAP50, and 2.0% in mAP50-95 over the YOLO11s baseline, with minimal increases in GFLOPs and parameters. RB-YOLO strikes an effective balance between detection performance and computational efficiency, making it well-suited for real-time maritime safety systems.