摘要
本文提出了一种基于多层PCA特征融合的深度学习模型来提取人脸遮挡特征﹒在传统主成分分析网络的基础上保留了输入图像的浅层结构特征,并对数据样本进行核(kernel)变换来提高模型的非线性拟合能力,然后通过融合各层的特征信息来丰富最终的遮挡特征表达,后期用随机森林算法进行分类器的训练﹒实验表明:在相同训练集和测试集下,本文提出的方法在网络训练效率和遮挡检测率上均要优于PCANet等人脸遮挡检测算法﹒
In this paper a deep learning model based on multi-layer PCA feature fusion is proposed to extract face occlusion features.Based on the analysis of the network on the retention of shallow structural features of the input image in the traditional principal component,and the nuclear data sample(kernel)transformation to improve the ability of nonlinear fitting model,and then by means of the fusion feature information of each layer to enrich the final block feature expression,the late random forest classifier training algorithm.Experiments show that under the same training set and test set,the proposed method is superior to PCANet and other face occlusion detection algorithms in terms of network training efficiency and occlusion detection rate.
作者
刘浩博
石跃祥
LIU Haobo;SHI Yuexiang(College of Information Engineering,Xiangtan University,Xiangtan,Hunan 411105,China)
出处
《湖南城市学院学报（自然科学版）》
CAS
2018年第1期43-47,共5页
Journal of Hunan City University:Natural Science
关键词
深度学习
遮挡检测
核变换
特征融合
随机森林
deep learning
occlusion detection
kernel transformation
feature fusion
random forest
