基于YOLO的多尺度并行人脸检测算法

Multi-scale parallel face detection algorithm based on YOLO

针对通用目标检测方法YOLO(you only look once)直接应用到人脸检测中存在召回率不够高、定位不够准确的问题,提出一种由密集到稀疏的多尺度并行的网络结构。通过不同尺度的网络检测不同尺寸的人脸,解决召回率不够高的问题,通过平均多尺度网络的检测结果解决定位不够准确的问题。引入中心损失函数,减小类内距离,进一步提高分类准确率。实验结果表明,在不同的数据集上,该方法的召回率及定位准确性相对于YOLO有所提高,检测精度接近主流方法的同时检测速度具有明显优势。

Aiming at the problem that the recall rate and the positioning are not satisfied based on the generic object detection algorithm YOLO(you only look once)directly,a multi-scale parallel network structure from dense to sparse for face detection was proposed.Different scale networks were used to detect different sizes of faces to solve the problem that the recall rate is not high enough,and the problem of inaccurate positioning was solved through the average multi-scale network detection results.The center loss function was introduced to reduce the intra-class distance and further improve the classification accuracy.Experimental results show that the recall rate and positioning of the proposed method are improved compared with YOLO,the detection accuracy is close to that of the mainstream method,and the detection speed is more superior on different datasets.