摘要
深度学习模型可以获得更具有鉴别力的人脸特征,提高人脸识别性能.因此,文中结合深度学习思想,提出多层次深度网络融合特征提取模型.在深度子空间基础上,采用"卷积-池化"网络结构,在降低特征维度的同时保留图像纹理信息,并且获得局部转换鲁棒性.同时,利用人脸标定算法获得人脸特征点,并以此划分人脸区域为5个局部人脸块.基于多层次分类策略,利用全局人脸训练全局网络,完成测试样本预分类.利用局部人脸块训练局部网络,在候选类别中完成最终分类.实验表明,结合局部特征与全局特征的模型可以取得较好的识别率,对光照、表情、姿态,遮挡等影响因素具有较好的鲁棒性,并且加入池化层及两步判别的算法可以有效提高识别率.
Discriminative facial features can be obtained by deep learning model. Therefore, combining the deep learning, a multi-level deep network extraction model for fusion feature is proposed. In the proposed model, the pooling layer is added after subspace mapping based on deep subspace model, so that the feature dimension is reduced with texture details preserving and local transformation robustness. Meanwhile, face region is divided into 5 parts according to facial feature point achieved by face alignment algorithm. Based on multi-level classification strategy, the global network is firstly trained using the whole face image to obtain five candidate labels for test sample. Then, the local face block is put into sub-network to obtain local representation and test samples are classified in the candidate labels. Experimental results show that the model combined with the local features and global features achieves better accuracy and robustness in the aspect of the illumination, expression, occlusion, etc. Moreover, adding pooling structure and the two-step discrimination algorithm effectively improve the recognition efficiency.
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2017年第5期448-455,共8页
Pattern Recognition and Artificial Intelligence
基金
国家自然科学基金项目(No.61071199)
河北省自然科学基金项目(No.F2016203422)资助~~
关键词
深度学习
子空间
池化
多层次融合
Deep Learning, Subspace, Pooling, Multi-level Fusion
