| 秦娥*,何佳瑶*,刘银伟*,朱娅妮**,李小薪*.基于稠密连接的通道混合式PCANet的低分辨率有遮挡人脸识别[J].高技术通讯(中文),2024,34(6):602~615 |
| 基于稠密连接的通道混合式PCANet的低分辨率有遮挡人脸识别 |
| Dense channel-hybrid PCANet for low-resolution and occluded face recognition |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2024. 06. 005 |
| 中文关键词: 有遮挡人脸识别； 主成分分析网络(PCANet)； 通道相关式卷积(CDC)； 稠密连接 |
| 英文关键词: face recognition with occlusion, principal component analysis network（PCANet）, channel-dependent convolution (CDC), dense connection |
| 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）
（**杭州电子科技大学计算机学院杭州 310018） | | 何佳瑶* | | | 刘银伟* | | | 朱娅妮** | | | 李小薪* | |
|
| 摘要点击次数: 257 |
| 全文下载次数: 236 |
| 中文摘要: |
| 针对低分辨率有遮挡人脸识别问题提出了基于稠密连接的通道混合式主成分分析网络(DCH-PCANet)。现有的PCANet模型的卷积层只使用了通道无关式卷积(CIC)。通道无关式卷积由于未考虑特征图在通道方向上的相关性，可以更好地凸显单个特征图的局部纹理特征，对于补偿因低分辨率、遮挡等因素导致的特征损失具有重要意义，但也会强化遮挡区域的特征，从而放大坏特征的影响范围；而通道相关式卷积(CDC)由于充分考虑了各特征图在通道方向上的相关性，可以较好地抑制坏特征的作用，形成较为稀疏的特征图。在PCANet中添加了基于通道相关式卷积的特征图提取分支，形成了通道混合式PCANet；并且引入了稠密连接，以充分利用低阶特征提升有遮挡图像识别的鲁棒性。针对如下4种数据集进行了实验：受控环境、真实遮挡和模拟低分辨率的人脸数据集（AR人脸数据集），非受控环境、真实遮挡和模拟低分辨率的人脸数据集（MFR2和PKU-Masked-Face），非受控环境、真实遮挡和真实低分辨率的人脸数据集（自建数据集）。实验结果表明，与现有方法相比，所提出的基于稠密连接的通道混合式PCANet具更好的遮挡鲁棒性和低分辨率鲁棒性，可以作为前沿方法的有效补充，提升其识别性能。 |
| 英文摘要: |
| A dense channel-hybrid PCANet (DCH-PCANet) is proposed to recognize low-resolution and occluded face images. Only channel-independent convolutions (CIC) are used in the convolutional layer of the existing principal component analysis network（PCANet） model. Since CIC does not consider the correlation of the feature maps in the channel direction, it can better highlight the local texture features of a single feature map, which is of great significance for compensating the feature loss caused by low resolution and occlusion. However, CIC will also strengthen the occlusion features, hence enlarging the influence range of bad features. The channel-dependent convolution (CDC) fully considers the correlation of all feature maps in the channel direction, which can better suppress the effect of bad features and form a sparse feature map. A CDC-based feature-map extraction branch is added to PCANet to form a channel-hybrid PCANet. And dense connections are also introduced to make full use of low-level features to improve the robustness of occluded image recognition. Experiments are conducted on the following four datasets: AR face dataset, where face images with real occlusions and simulated low-resolutions are acquired in controlled environment; MFR2 and PKU-Masked-Face, where face images with real occlusions and simulated low-resolutions are acquired in uncontrolled environment; our own dataset, where face images with real occlusion and real low-resolution are acquired in uncontrolled environment. Experimental results show that compared with the existing methods, the proposed DCH-PCANet has better occlusion and low-resolution robustness, which can be used as an effective supplement to the cutting-edge methods to improve their recognition performance. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
