基于锚框密集化与度量学习的多尺度人脸检测研究

Study on Multi-scale Face Detection Based on Anchor Box Densification and Metric Learning

本文针对基于单次多盒检测算法在检测小目标时性能不佳的问题,提出一种改进异构单次多盒网络的上下文语义辅助增强特征的人脸检测算法,包括借助于辅助的全卷积网络而构建的特征增强模块、度量函数设计以及针对小脸的锚框匹配密集化方法。首先,将固定大小的单比例图像作为输入,并以完全卷积的方式在两个异构CNN上输出按比例大小的特征图,特征图的增强模块通过自上而下和横向连接两个异构网络的原始特征图,得到上下文语义增强的特征金字塔;其次,根据异构网络的特征提取的互补性来等设计比例间隔并密集分布的锚框,通过改进的度量损失来端到端地训练整个网络。最后,对公开的WIDER FACE和FDDB数据集进行了实验,结果表明该方案获得了较好的人脸检测性能。

Aiming at the problem of the poor performance of the single-shot multi-box detection algorithm in detecting small targets,this paper proposes a face detection algorithm based on an improved heterogeneous single-shot multi-box network with contextual semantic assisted feature enhancement,which include feature enhancement module constructed by the convolutional network,the metric function design,and the anchor frame matching densification method for small faces.First,a fixed-size single-scale image was used as input,and a proportional-size feature map was output on two heterogeneous CNNs in a fully convolutional manner.The enhancement module of the feature map connected the two different images of original feature map from top to bottom and horizontally,then a feature pyramid with enhanced context semantics was obtained;secondly,according to the complementarity of the feature extraction of heterogeneous networks,proportionally spaced and densely distributed anchor frames were designed,and entire network were trained end-to-end through improved metric loss.Finally,experiments were carried out on the public WIDER FACE and FDDB data sets,and the results showed that the scheme has achieved better face detection performance.