In the realm of unmanned surface vehicle applications, detecting small, low-resolution targets on water surfaces poses significant challenges due to limited pixel information and blurred details. This study introduces a novel detection model, termed Hybrid Feature Prominence and Adaptive Intersection over Union Ratio Fusion YOLO (HFA-YOLO), designed specifically for this purpose. By integrating a Channel-Space Hybrid Feature Prominence Module (CSHM) into the backbone network, the model enhances fine-grained features such as edges and textures, preserving small target details during multi-scale feature fusion. Additionally, an adaptive IoU computation mechanism is introduced to jointly assess positional deviations and shape similarity between predicted and ground truth boxes, improving localization accuracy and robustness. Experimental results on the FloW-Img and WSODD datasets demonstrate that HFA-YOLO achieves mean average precision (mAP) scores of 79.8% and 80.6%, respectively, with an inference speed of 110 frames per second (FPS), outperforming mainstream detection algorithms. This approach satisfies the demands of high-precision, real-time small object detection, offering valuable insights for environmental monitoring, water rescue, and surface garbage detection applications.
Liu et al. (Wed,) studied this question.