Abstract: To address insufficient multi-scale feature fusion and edge detail loss in Unmanned Aerial Vehicle (UAV) images, an improved YOLO12n-based Edge and Frequency Aware Network (EFA-Net) was proposed. Firstly, an edge-aware feature extraction module, Edge-Aware Stem (EAStem), was designed to enhance fine-grained edge and low-level structural feature extraction. Secondly, a frequency-aware cross-stage partial module, Frequency-Aware C2f (FAC2f), was proposed to enhance the model's perception and recognition of complex distresses by efficiently fusing spatial and frequency domain features. Finally, a Multi-source collaborative aggregation enhanced Feature Pyramid Network (MFPN) was constructed to mitigate semantic information loss and optimize cross-layer feature alignment. Experimental results on the UAV-PDD2023 dataset demonstrate that EFA-Net achieves mAP50 and mAP50-95 of 80.66% and 55.54%, respectively. The model attains superior detection accuracy while maintaining low computational complexity.

Key words: Unmanned Aerial Vehicle (UAV) image, pavement distress, edge awareness, frequency feature fusion, Feature Pyramid Network (FPN)

摘要： 针对无人机（UAV）图像中多尺度特征融合不足及边缘细节信息丢失等问题，提出了一种基于YOLO12n改进的边缘频率感知网络(EFA-Net)。首先，设计边缘感知特征提取模块EAStem（Edge-Aware Stem），增强对细粒度边缘与低层结构特征的提取能力；其次，提出频率感知跨阶段部分模块FAC2f（Frequency-Aware C2f），通过高效融合空间域与频域特征，增强模型对复杂病害的特征感知与识别能力；最后，构建多源特征协同聚合增强型特征金字塔网络(MFPN)，有效缓解语义信息丢失，并提升跨层特征的对齐能力。在UAV-PDD2023数据集上的实验结果表明，EFA-Net的mAP50与mAP50-95分别达到80.66%与55.54%，在保持较低计算复杂度的同时实现了更优的检测精度。

关键词: 无人机图像, 路面病害, 边缘感知, 频率特征融合, 特征金字塔网络