Abstract: To address the problems of densely distributed small targets, weak feature information, and insufficient multi-scale feature fusion in UAV-view images, an improved detection model, Frequency-enhanced Progressive Metric-optimized YOLO (FPM-YOLO), was developed based on YOLOv8. First, the detection structure was reconstructed by removing the P5 detector head and introducing the P2 detector head, making the model focus more on high-resolution small target features. This reduces its computational cost and enhances its ability to detect small targets. Second, the PRN (Progressive Refinement Neck) structure was introduced into the neck network, and the effect of multi-scale feature fusion was enhanced through backbone feature reuse and progressive fusion, thereby improving the reuse of low-level features and high-level semantic information. Then, a Gaussian Combined Distance (GCD) metric optimization strategy was introduced and applied to both bounding box regression loss and label assignment to improve the matching quality and localization accuracy of small target bounding boxes. Finally, an EHFM (Edge-guided High-frequency Enhancement Module) was constructed by combining edge guidance with high-frequency enhancement in the frequency domain to strengthen the contour and texture features of small targets. Experimental results show that, compared with the baseline model, precision, recall, mAP50, and mAP50-95 are improved by 6.8, 8.9, 10.0, and 6.6 percentage points, respectively, on the VisDrone2019 dataset, which verifies the effectiveness of the proposed model for Unmanned Aerial Vehicle (UAV) small target detection tasks.

Key words: YOLOv8, small target detection, progressive fusion, multi-scale, drone aerial photography

摘要： 针对目前无人机视角下的图像存在小目标分布密集、特征信息弱且多尺度特征融合不足等问题，提出一种基于YOLOv8改进的检测模型FPM-YOLO（Frequency-enhanced Progressive Metric-optimized YOLO）。首先，对检测结构进行重构，删除P5检测头并引入P2检测头，使模型更加关注高分辨率小目标特征；其次，在颈部网络中使用PRN（Progressive Refinement Neck）结构进行改进，通过骨干特征复用与渐进式融合加强多尺度特征融合效果，提升低层特征与高层语义信息的复用率；再次，引入GCD（Gaussian Combined Distance）度量优化策略，将它同时用于边框回归损失与标签分配，以提高小目标边框匹配质量和定位精度；最后，将边缘引导与频域高频增强相结合构造EHFM（Edge-guided High-Frequency enhancement Module），增强小目标轮廓与纹理等特征信息。实验结果表明，改进模型在VisDrone2019数据集上的精确率、召回率、mAP50和mAP50-95相较于基准模型分别提升了6.8、8.9、10.0、6.6个百分点，验证了所提模型在无人机（UAV）小目标检测任务中的有效性。

关键词: YOLOv8, 小目标检测, 渐进式融合, 多尺度, 无人机航拍