UAV maritime small object detection via frequency-spatial dual-domain synergistic perception and multi-granularity feature fusion

In UAV-based maritime search-and-rescue (SAR) and surveillance scenarios, small object detection remains challenging because distant targets occupy very few pixels and are easily affected by wave clutter, whitecaps, and sun-glint interference. To address these issues, this study proposes EMT-YOLO, an efficient detector for maritime small object detection that integrates frequency-spatial dual-domain perception with multi-granularity feature fusion. Specifically, a Multi-Scale Adaptive Wavelet Transform Convolution (MSAWTConv) module is introduced to enhance frequency-aware feature extraction and suppress clutter-dominated responses. DySample is adopted to improve content-aware feature reconstruction during cross-scale fusion. In addition, a Multi-Scale Convolutional Attention (MSCA) mechanism is incorporated to strengthen spatial focusing on weak target regions under complex maritime backgrounds. The detection pyramid is further reconstructed by introducing a high-resolution P2 detection branch and removing the P5 detection branch from the final prediction head, so as to better match the scale distribution of maritime small objects. Experimental results on the SeaDroneSee dataset show that EMT-YOLO improves mAP@50 by 14.2 percentage points and Recall by 14.9 percentage points relative to YOLO11n, while reducing the number of parameters by 17.4%. These results indicate that EMT-YOLO provides a favorable accuracy-efficiency trade-off for maritime small object detection.