MarineSwin-DETR: A Frequency-Adaptive and Dynamic Histogram Attention Fusion Framework for Marine Small Object Detection

With the rapid development of remote sensing technology, the demand for accurate detection of small targets on the sea surface is increasingly urgent in the field of maritime monitoring. However, complex sea conditions such as strong solar flares, dynamic reflection of waves, and non-uniform illumination lead to three core challenges for traditional detection methods, namely, the lack of small target features, pseudo-target interference, and dynamic deformation. To this end, this study proposes the MarineSwin-DETR innovation framework, which mainly includes the following two innovations. Frequency domain Adaptive Dilated Convolution (FADC) module: By dynamically adjusting the dilation rate and kernel weight, the high-frequency details of multi-scale features are preserved and the low-frequency background is suppressed. Dynamic Range Histogram Self-attention (DHSA) mechanism is used to reconstruct the attention distribution based on the intensity bucketing strategy, which effectively solves the contrast degradation problem under high dynamic illumination. Experiments on the SeaDronesSee dataset show that the proposed model achieves 84.6% mAP@0.5 at a computational cost of 55.0 GFLOPs, which is 2.2 percentage points higher than the baseline model RT-DETR r18, and the detection accuracy of small objects such as life-saving equipment is improved by 9.5%-10.1%. This result provides a solution with both high accuracy and low power consumption for real-time target detection in complex sea conditions.