基于多任务学习的多姿态人脸重建与识别

发布时间：2018-04-17 03:34

本文选题：多任务学习 + 姿态恢复　；参考：《计算机应用》2017年03期

【摘要】：针对当前人脸识别中姿态变化会影响识别性能,以及姿态恢复过程中脸部局部细节信息容易丢失的问题,提出一种基于多任务学习的多姿态人脸重建与识别方法——多任务学习堆叠自编码器(Mt LSAE)。该方法通过运用多任务学习机制,联合考虑人脸姿态恢复和脸部局部细节信息保留这两个相关的任务,在步进逐层恢复正面人脸姿态的同时,引入非负约束稀疏自编码器,使得非负约束稀疏自编码器能够学习到人脸部的部分特征;其次在姿态恢复和局部信息保留两个任务之间通过共享参数的方式来学习整个网络框架;最后将重建出来的正脸图像通过Fisherface进行降维并提取具有判别信息的特征,并用最近邻分类器来识别。实验结果表明,Mt LSAE方法获得了较好的姿态重建质量,保留的局部纹理信息清晰,而且与局部Gabor二值模式(LGBP)、基于视角的主动外观模型(VAAM)以及堆叠步进自编码器(SPAE)等经典方法相比,识别率性能得以提升。
[Abstract]:In view of the problem that the change of pose in current face recognition will affect the performance of face recognition, and the local details of face are easily lost in the process of attitude recovery,A multi-pose face reconstruction and recognition method based on multi-task learning is proposed, which is multi-task learning stacked self-encoder.By using multi-task learning mechanism and considering the two related tasks of facial pose restoration and facial local detail information retention, this method introduces a non-negative constraint sparse self-encoder to restore the frontal face pose layer by layer at the same time.The non-negative constraint sparse self-encoder can learn some features of human face; secondly, the whole network framework can be learned by sharing parameters between the two tasks: attitude recovery and local information retention.Finally, the reconstructed face image is reduced by Fisherface and the features with discriminant information are extracted, and recognized by nearest neighbor classifier.The experimental results show that the modified Mt LSAE method has better attitude reconstruction quality and retains clear local texture information, and compared with the classical methods such as local Gabor binary mode (LGBPU), visual angle based active appearance model (VAAM) and stacking step self-encoder (SPAE).The recognition rate is improved.
【作者单位】：认知无线电与信息处理省部共建教育部重点实验室(桂林电子科技大学);桂林电子科技大学信息与通信学院;
【基金】：国家自然科学基金资助项目(61362021,61661017) 广西自然科学基金资助项目(2013GXNSFDA019030,2014GXNSFDA118035) 广西科技创新能力与条件建设计划项目(桂科能1598025-21) 桂林科技开发项目(20150103-6)~~
【分类号】：TP391.41

【相似文献】