HL-YOLO: high–low frequency feature fusion for maritime small target detection
Abstract
Maritime small-object detection from UAV imagery is a fundamental perception task for coastal surveillance and search-and-rescue, aiming to localize and classify tiny targets such as swimmers and vessels in complex sea-surface scenes. FFCA-YOLO provides a strong YOLOv5-based baseline for small objects, yet its hierarchical fusion still lacks an explicit mechanism to jointly exploit fine-grained spatial details and high-level semantics under wave clutter and specular reflections. We propose HL-YOLO with a unified Spatial-Perceptual High–Low Feature Fusion (SPHLFF) module that integrates three components: Spatial Perception Decoupled Enhancement (SPDE) for separate high/low-frequency enhancement, Cross-Level Mutual-Guided Weighting Mechanism (CMWM) for bidirectional recalibration, and Top-K Sparse Self-Attention (TKSA) for multi-scale sparse attention with background-clutter suppression. On S-ODv2, HL-YOLO achieves 0.777 AP50 and 0.423 mAP@0.5:0.95 with 7.23M parameters and 216 FPS, showing a favorable accuracy–efficiency trade-off for real-time maritime surveillance.