改进YOLO的密集小尺度人脸检测方法  

Improved YOLO dense small-scale face detection method

作　　者：邓珍荣 白善今[1] 马富欣 黄文明 DENG Zhen-rong;BAI Shan-jin;MA Fu-xin;HUANG Wen-ming(College of Computer and Information Security,Guilin University of Electronic Technology,Guilin 541004,China;Guangxi Colleges and Universities Keys Laboratory of Cloud Computing and Complex Systems,Guilin University of Electronic Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学计算机与信息安全学院,广西桂林541004 [2]桂林电子科技大学广西高校云计算与复杂系统重点实验室,广西桂林541004

出　　处：《计算机工程与设计》2020年第3期874-879,共6页Computer Engineering and Design

基　　金：广西自然科学基金项目(2018GXNSFAA138132);桂林电子科技大学研究生教育创新计划基金项目(2018YJCX55)。

摘　　要：为解决密集小尺度人脸检测精度低的问题,提出一种改进YOLO的密集小尺度人脸检测方法。使用目标框与真实框的面积交并比作为距离损失函数对传统的k-means聚类算法进行改进,结合小尺度人脸目标占比小且长宽比例接近1的特点,对候选框进行聚类,筛选合适的尺度数量;在不同层级的特征图进行细粒度的特征融合,对感受野较小的浅层特征进行空间降维通道升维后,与感受野较大的深层特征进行融合,提高对小尺度人脸特征的表示能力;结合聚类框,调整预测层的宽度和深度,形成适用于检测密集小尺度人脸网络结构。在WIDER FACE人脸检测数据库上进行实验的结果表明,该方法对密集小尺度人脸的检测精度有明显提高。To solve the problem of low precision of dense small-scale face detection,an improved YOLO dense small-scale face detection method was proposed.The traditional K-means clustering algorithm was improved using the area intersection ratio of the target frame and the real frame as the distance loss function.Combined with the characteristics that the small-scale face target proportion is small and the length width ratio closes to 1,the appropriate number of scales of the candidate frame was selected by clustering.Fine-grained feature fusion was carried out in different levels of feature maps,and the shallow features with smaller receptive fields were spatially reduced and the channels were upgraded and then fused with the deep features with larger receptive fields,to improve the ability to express small-scale facial features.Combined with the clustering box,the width and depth of the prediction layer were adjusted to form a small-scale face network structure suitable for detecting.The results of experiments on the WIDER FACE faces detection database show that the detection accuracy of the method for dense small-scale faces is signi-ficantly improved.

关 键 词：人脸检测 YOLO网络 聚类 密集小尺度人脸 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]