一种改进的多任务级联卷积神经网络人脸检测算法

An Improved Multi Task Cascaded Convolutional Neural Network Face Detection Algorithm

基于卷积神经网络的人脸检测方法检测准确率高,应用广泛,但卷积神经网络一般结构复杂,运算量大,在边缘计算设备运行时实效性不高,为提高实效性,本文提出一种改进的多任务卷积神经网络人脸检测算法。该检测方法基于MTCNN实现,使用MobileNet替换MTCNN三层子网络中的大部分卷积核,并减少R-Net、O-Net网络中的卷积核数量;改进NMS算法,降低被抑制候选框的置信度,在最后时统一删除,并抑制同一人脸周围分数最高候选框之外的其他候选框。选用Celeb A和WIDER FACE数据集按照7:3划分进行训练和验证,准确率达到了98.03%,实时性较好。该改进策略大幅降低了原网络的参数量,实现了轻量化改进,保留精度更高的回归窗口,缓解了漏检问题,检测率高,能满足边缘算力设备人脸实时检测的需求。

Face detection method based on convolutional neural network has high detection accuracy and wide application,but the convolutional neural network has a complex structure,large amount of calculation,and low effectiveness when running on edge computing devices.In order to improve the real-time detection,an improved multi-task convolutional neural network face detection algorithm is proposed.The detection method is realized based on MTCNN,using MobileNet to replace most of the convolution kernels in the three layer sub-network of MTCNN,and reducing the number of convolution kernels in R-Net and O-Net networks;improve the NMS algorithm,reduce the confidence of the suppressed candidate box,delete it uniformly at the last time,and suppress other candidate boxes except the highest score candidate box around the same face.When detected on edge computing devices,Celeb A and WIDER FACE datasets are selected for training and verification according to the 7:3 division,with an accuracy of 98.03%and good real-time performance.The improved strategy greatly reduces the amount of parameters of the original network,realizes lightweight improvement,retains the regression window with higher accuracy,alleviates the missed detection problem,has a high detection rate,and can meet the demand of real-time face detection on edge computing devices.