基于回归置信度的多尺度精确人脸检测

在大多数受限情况下人脸检测已经有了许多有效方案,但对于人脸尺度变化极大、小人脸,以及模糊、遮挡、光照等非受限环境的人脸检测问题,仍面临更多挑战。针对以上问题,提出一种多尺度卷积神经网络模型。在R-FCN网络的基础上进行改进,以多尺度特征替代单一特征,使网络对多尺度信息更加敏感,在预测阶段同时输出分类置信度与回归置信度,改进非极大值抑制(non-maximum suppression,NMS)算法,提出基于回归置信度的NMS算法。在WIDER FACE数据集上训练模型,在FDDB与WIDER FACE人脸评测库进行实验,实验结果表明,召回率、准确率等指标均优于其它人脸检测算法。

Multi-scale accurate face detection based on regression confidence

Numerous accurate and efficient face detection methods have been proposed for most constrained scenarios,but works focus on detecting faces in uncontrolled scenarios,which is more challenging due to faces with large scale changes,tiny faces,and faces under conditions of blurring,occlusion and illumination.In the view of these problems,a multi-scale convolution neural network was proposed.The model was improved on the basis of R-FCN network,replacing the original single feature with multi-layer features,which made the network more sensitive to multi-scale information,getting the classification score and regression score in the forecast stage simultaneously,changing the original NMS algorithm and proposing regression score based NMS algorithm.The model was trained on WIDER FACE dataset,and experienced on FDDB and WIDER FACE benchmark.The results show that the proposed algorithm can make state of art performance.