Anti-interference multi-scale detection algorithm for unmanned surface vehicles based on dynamic feature enhancement

Target detection for unmanned surface vehicles (USVs) in complex marine environments suffers from scale-varying targets and environmental interference, leading to high false positive and missed detection rates. To address these limitations, this paper proposes YOLO-MKV (YOLO-Marine Knowledge Vision)—a dynamic feature-enhanced object detection algorithm. First, a novel Multi-Scale Adaptive Kernel Convolution (MSAK) module is combined with a lightweight Mixed Local-Channel Attention (MLCA) mechanism. This innovative integration leverages dynamically tuned multi-scale kernels for the feature-adaptive extraction of variably-sized objects, while utilizing MLCA to seamlessly fuse local spatial details with global contextual statistics across channel dimensions, improving multi-scale representation and shape adaptability with low computational cost. In order to enhance the extraction of subtle features and the relationships within the global context, an additional high-resolution P2 detection head is fused with the C3k2-Vision Transformer module. Furthermore, a dynamic non-monotonic focusing mechanism based on the WIoU v3 loss function is employed to apply an adaptive gradient gain strategy that prevents overfitting to high-quality anchor boxes and suppresses harmful gradients from outliers, ensuring precise bounding box localization. Extensive experiments on the WSODD dataset demonstrate that YOLO-MKV achieves 82.0% mAP@0.5 and 47.3% mAP@0.5:0.95, which are 6.5% and 5.4% higher than the baseline model, respectively. With only 2.67 M parameters and 11.0 GFLOPs, our method achieves competitive performance among the compared detectors such as YOLOv8s and YOLOv11s.