基于区域渐进校准网络的人脸检测与定位

Face detection and localization on region-base progressive calibration network

为解决角度变化下的人脸检测中存在参数量大及角度幅度变量小的问题,提出区域渐进校准网络用于任意平面角度的人脸检测,通过级联网络结构降低角度变化、提升网络运行速度。采用区域生成网络产生高质量的候选区域,构造渐进校准网络,逐步缩小面部平面角度变化范围,同时由粗到细地对候选区域执行面部检测。其中,特征提取的中间层融合参数量较少时,更好地表示了面部特征,调整锚的设置解决小尺度面部问题。在角度增强的FDDB(face detection data set and benchmark)数据集与WIDER FACE数据集上的实验结果表明,提出的方法分别取得了89.1%与90.4%的平均召回率,准确度高于快速区域卷积神经网络(Faster RCNN),且运行速度更快。在实际项目中使用该算法,验证了该方法的有效性及可行性。

In order to solve the problem of large parameter quantity and small angle amplitude variable in face detection under angle change,this paper proposes a region-based progressive calibration network for face detection at any plane angle,which reduces the angle change and improves the network running speed through the cascade network structure.Specifically,the region generation network is first used to generate a high quality candidate region,and then the progressive calibration network is constructed to gradually reduce the range of the face plane angle variation,while performing face detection on the candidate region from coarse to fine.Among them,when the number of middle-layer fusion parameters for feature extraction is small,the facial features are better represented,and the anchor setting is adjusted to solve the small-scale facial problem The experimental results on the angle-enhanced FDDB(face detection data set and benchmark)and the WIDER FACE dataset show that the proposed method achieves an average recall rate of 89.1%and 90.4%,respectively,achieving more accuracy than that of Faster RCNN and running fast.The effectiveness and feasibility of this method are verified by using this algorithm in a practical project.