Abstract: Aiming at the problem of low face detection accuracy caused by changes in lighting, low resolution, posture and expression, and the generalization of algorithms caused by most face detection algorithms using a single convolutional neural network to extract features, a three-layer convolutional neural network structure consisting of shallow and deep cascade is proposed. The face candidate region is quickly located by the full convolutional neural network. Then the depth neural network is used to extract the face robustness feature, and the candidate region is further classified and verified. The joint regression face method is used to determine the final face position and improve the detection accuracy. At the same time, the commonly used non-maximum value suppression method is improved by weighting the reduction score, and the missed detection problem caused by the overlapping of adjacent faces is solved. The experimental results show that the model is robust to the above-mentioned factors that cause low face detection accuracy, and it has high accuracy and running speed in FDDB dataset. The improved non-maximum suppression algorithm also has a certain improvement on the test accuracy of FDDB.

Key words: face detection, full convolutional network, joint regression

摘要： 针对由于光照、分辨率、姿态和表情等因素变化引起的人脸检测准确性不高的问题和大多人脸检测算法使用单一的卷积神经网络去提取特征引起的算法的泛化能力变弱的问题，提出了三层由浅及深的级联的卷积神经网络结构。通过全卷积神经网络快速定位人脸候选区域，采用深度神经网络提取人脸鲁棒性特征，对候选区域进一步分类验证，并用联合回归的方法确定最终人脸位置，提高检测精度。同时通过加权降低得分改进常用的非极大值抑制的方法，解决了由于相邻人脸高度重叠引起的漏检问题。实验结果表明，该模型对上述引起人脸检测准确率不高的因素具有较好的鲁棒性，并且在FDDB数据集上有着较高的准确率和运行速度。改进后的非极大值抑制算法对在FDDB的测试准确率也有一定的提升。

关键词: 人脸检测, 全卷积网络, 联合回归