复杂场景下多姿态人脸知识蒸馏识别方法  

Multi-pose Face Knowledge Distillation and Recognition Method in Complex Scene

作　　者：孙兰兰[1] Sun Lanlan(Tongcheng Teachers College,Tongcheng,Anhui 231400,China)

机构地区：[1]桐城师范高等专科学校,安徽桐城231400

出　　处：《黑龙江工业学院学报（综合版）》2022年第9期92-97,共6页Journal of Heilongjiang University of Technology(Comprehensive Edition)

基　　金：安徽高校自然研究重点项目“基于知识蒸馏型卷积神经网络的复杂场景下人脸识别研究”(项目编号:KJ2929A0892)。

摘　　要：目前使用的多姿态人脸识别方法受到人脸旋转角度影响,出现人脸表情识别结果与实际表情不一致的问题。为了避免该问题的产生,研究了复杂场景下多姿态人脸知识蒸馏识别方法。通过采集人脸多姿态采样样本图像,得到人脸在复杂场景下旋转角度值,即偏航角、俯仰角、翻滚角;通过增强随机信号构建人脸数据训练网络,依次训练输入样本,利用分类处理获取分类层输出结果;采用归一化处理输出层次特征,计算两层之间特征距离,构建知识蒸馏损失函数。根据复杂场景偏航角、俯仰角、翻滚角区间分类概率,计算角度回归损失。根据角间隔区间概率输出结果,构建人脸知识蒸馏模型。使用三维地标检测方法提取人脸关键点,结合径向基函数重构人脸模型。计算偏航角方向、俯仰角方向、翻滚角方向的人脸姿态参量,结合蒸馏点损失自适应调整结果,识别多姿态人脸。实验结果表明,虽然使用该方法出现了与实际参数最大误差为2mm的情况,但是并不影响表情识别结果,能够得到与实际表情一致的多姿态人脸。The current multi-pose face recognition methods are affected by the face rotation angle,and the result of facial expression recognition is inconsistent with the actual expression.In order to avoid this problem,the knowledge distillation and recognition method for multi-pose faces in complex scenes is studied.Face multi-pose sample images are collected to obtain the values of face rotation angle in complex scenes,namely yaw angle,pitch angle and roll angle;The face data training network is constructed by enhancing the random signal,the input samples are trained in turn,and the output results of the classification layer are obtained by classification processing;Normalization is used to process the output hierarchical features,the feature distance between the two layers is calculated,and the knowledge distillation loss function is constructed.The angle regression loss is calculated according to the interval classification probability of yaw angle,pitch angle and roll angle in complex scenes.According to the probability output results of angular interval,the face knowledge distillation model is constructed.The 3D landmark detection method is used to extract the key points of the face,and the radial basis function is used to reconstruct the face model.The face pose parameters of yaw angle direction,pitch angle direction and roll angle direction are calculated.Combined with the adaptive adjustment results of distillation point loss,the multi-pose face is recognized.The experimental results show that although the maximum error between the proposed method and the actual parameters is 2mm,it does not affect the expression recognition results,and the multi-pose face consistent with the actual expression can be obtained.

关 键 词：复杂场景 多姿态 人脸识别 知识蒸馏 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]