一种多层特征融合的人脸检测方法

由于姿态、光照、尺度等原因，卷积神经网络需要学习出具有强判别力的特征才能应对复杂场景下的人脸检测问题。受卷积神经网络中特定特征层感受野大小限制，单独一层的特征无法应对多姿态多尺度的人脸，为此提出了串联不同大小感受野的多层特征融合方法用于检测多元化的人脸；同时，通过引入加权降低得分的方法，改进了目前常用的非极大值抑制算法，用于处理由于遮挡造成的相邻人脸的漏检问题。在FDDB和WiderFace两个数据集上的实验结果显示，文中提出的多层特征融合方法能显著提升检测结果，改进后的非极大值抑制算法能够提升相邻人脸之间的检测准确率。

Face detection method fusing multi-layer features

To address the issues of pose, lighting variation, and scales, convolutional neural networks (CNNs) need to learn features with strong discrimination handle the face detection problem in complex scenes. Owing to the size limitations of the specific feature layer’s receptive field in convolutional neural networks, the features computed from a single layer of the CNNs are incapable of dealing with faces in multi poses and multi scales. Therefore, a multi-layer feature fusion method that is realized by fusing the different sizes of receptive fields is proposed to detect diversified faces. Moreover, via introducing the method of weighted score decrease, the present usual non-maximum suppression algorithm was improved to deal with the detection omission of neighboring faces caused by shielding. The experiment results with the FDDB and WiderFace datasets demonstrated that the fusion method proposed in this study can significantly boost detection performance, while the improved non-maximum suppression algorithm can increase the detection accuracy between neighboring faces.