Abstract: High-precision face detection algorithms and lightweight counterparts suffer from opposing limitations in UAV deployment: the former are computationally prohibitive, whereas the latter lack the accuracy required for dense crowd scenarios.To address these issues,in this paper YOLO ECEL, a lightweight multi-scale feature enhancement algorithm for dense crowd face detection in aerial images, was proposed based on YOLO11n. First, a lightweight backbone EfficientNetV2 E embedded with Efficient Multi-Scale Attention (EMA) was constructed to enhance multi scale face feature extraction. Then, the CCFM Dy neck network was built by integrating Cross-Scale Feature Fusion Module (CCFM) cross scale feature fusion with DySample dynamic upsampling to preserve edge details of small targets. Third an Efficient Multi Receptive Field Feature Enhancement (EfficientMRFFE) module was designed to strengthen feature representation for dense overlapping faces. Finally, the Lightweight Shared Detail-Enhanced Convolutional Detection Head (LSDECD Head) was adopted to effectively reduce computational redundancy. Experimental results on the self built aerial dataset UAV DenseFace show that, compared with the baseline algorithm, YOLO ECEL improves mAP50 and mAP50:95 by 0.4 and 0.3 percentage points respectively, and reduces the number of parameters, computational cost and model size by 42.3%, 17.2% and 36.5%. Thus, it achieves an efficient and lightweight design while maintaining detection accuracy, and can better support UAV security tasks in dense crowd scenarios.

Key words: Public safety, UAV, Dense face detection, Lightweight object detection, YOLO-ECEL

摘要： 针对高精度人脸检测算法算力开销大、机载部署困难，传统轻量算法在航拍密集人群场景检测精度不足的问题，本文基于YOLO11n提出一种轻量多尺度特征增强的航拍密集人群人脸检测算法YOLO-ECEL。首先，构建嵌入高效多尺度注意力（EMA）的EfficientNetV2-E轻量化主干，提升多尺度人脸特征提取能力；其次，结合跨尺度特征融合模块（CCFM）与DySample动态上采样，构建CCFM-Dy颈部网络，在降低计算开销的同时保留小目标边缘细节；再次，设计多感受野特征增强模块（EfficientMRFFE），强化密集重叠人脸特征表征能力；最后，采用轻量共享细节增强卷积检测头（LSDECD-Head），有效削减运算冗余。在自建航拍数据集UAV-DenseFace上的实验结果表明：相较于基线算法，YOLO-ECEL的mAP50和mAP50:95分别提升0.4和0.3个百分点，参数量、运算量和模型体积分别降低42.3%、17.2%和36.5%，在保证检测精度的前提下实现了高效轻量化，可更好支撑密集人群场景下的无人机安防任务。

关键词: 公共安全, 无人机, 人脸检测, 密集人群, 轻量化网络