| 秦娥*,卢天宇*,李卫锋**,刘银伟*,朱娅妮***,李小薪*.基于扩展的PCANet的有遮挡人脸识别方法[J].高技术通讯(中文),2025,35(2):134~144 |
| 基于扩展的PCANet的有遮挡人脸识别方法 |
| Occluded face recognition via extended PCANet |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 02. 003 |
| 中文关键词: 有遮挡人脸识别； 主成分分析模型； 稠密连接； 稠密编码； 滤波器多样性 |
| 英文关键词: face recognition with occlusion, principal component analysis network (PCANet), dense connections, dense encoding, filter diversity |
| 基金项目: |
| 作者 | 单位 | | 秦娥* | （*浙江工业大学计算机科学与技术学院杭州 310023）
（**浙江华通云数据科技有限公司杭州 310013）
（***杭州电子科技大学计算机学院杭州 310018） | | 卢天宇* | | | 李卫锋** | | | 刘银伟* | | | 朱娅妮*** | | | 李小薪* | |
|
| 摘要点击次数: 11 |
| 全文下载次数: 11 |
| 中文摘要: |
| 针对有遮挡人脸识别问题，本文将现有的卷积神经网络（convolutional neural networks，CNN）模型与主成分分析模型(principal component analysis network，PCANet)相结合，提出了扩展的PCANet(extended PCANet，xPCANet)模型。为了有效消除人脸图像中可能包含的遮挡信息造成的影响，通常需要充分利用网络的底层特征并构建尽可能丰富的特征。PCANet的2个不足在于：(1)由于正交性约束，各卷积层的滤波器高度相似，降低了滤波器响应的多样性；(2)在进行模式图编码时，对特征图进行了二值化处理，并采用了跨度较大的编码方式，从而丢弃了过多的信息。为了使PCANet能够更好地适配现有的CNN模型，在PCANet模型中引入了2个稠密连接：(1)在各卷积层之间引入了稠密连接，以充分利用底层卷积层提取的特征，并尽可能降低卷积层之间滤波器的相似性；(2)在PCANet的模式图编码阶段引入了加权稠密编码，以充分利用卷积层输出的特征生成更多的模式图。这2种稠密连接或编码方案都会进一步提升PCANet最终输出的柱状图特征的维度，并生成更为丰富的特征。在受控环境和有真实遮挡的人脸数据集（增强现实（AR）人脸数据集）、非受控环境和有模拟遮挡的数据集（LFW和CFP）、非受控环境和有真实遮挡的数据集（MFR2和PKU-Masked-Face）上的实验结果表明，所提扩展的PCANet模型能够有效处理实物遮挡和因光照引发的遮挡，也可以作为前沿方法的有效补充，提升前沿方法的遮挡鲁棒性。 |
| 英文摘要: |
| To solve the problem of occluded face recognition, an extended principal component analysis network(xPCANet) model is proposed by combining the existing convolutional neural networks (CNN) model with PCANet. In order to effectively eliminate the influence caused by facial occlusions, it is usually necessary to make full use of the low-level features of the network and build rich features. Two disadvantages of PCANet are: (1) due to the orthogonality constraint, the filters of each convolution layer are highly similar, which reduces the diversity of the filter response; (2) to get the pattern maps, the feature maps are binarized, and the encoding method with a large stride is adopted, so that many useful features are discarded. In order to make PCANet better fit into the existing CNN models, two dense connections are introduced into the PCANet model: (1) the dense connections introduced between convolutional layers are used to make full use of the features extracted by the low-level convolutional layers, and reduce the similarity of filters between convolutional layers as much as possible; (2) in the pattern-map encoding stage, weighted dense encoding is introduced to make full use of the features produced by the convolutional layers to generate more pattern maps. These two dense connections enhance the dimension of the final output of PCANet histogram features and generate richer features. Experiments on face datasets (AR face dataset) with real occlusions acquired in the controlled environment, on face datasets (LFW and CFP) with synthetic occlusions acquired in the uncontrolled environment, and on face datasets (MFR2 and PKU-Masked Face) with real occlusions acquired in the uncontrolled environment show that, compared with existing methods the proposed xPCANet can effectively deal with physical occlusions and illumination-caused occlusions, and can also be an effective supplement to the cutting-edge methods to improve their robustness against occlusions. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
