摘 要：提出了一种新的深度子空间聚类方法，使用了卷积自编码器将输入图像转换为位于线性子空间上的表示。通过结合自编码器提取的低阶和高阶信息来促进特征学习过程，在编码器的不同层级生成多组自我表示和信息表示。将得到的多级信息融合得到统一的系数矩阵并用于后续的聚类。通过多组实验验证了上述创新的有效性，在三个经典数据集：Coil20，ORL 和Extended Yale B 上，聚类精度分别达到 95.38%、87.25% 以及 97.58%。相较于其他主流方法，能有效提高聚类准确性，并具有较强的鲁棒性。
关键词：子空间聚类；多级结构；自编码器
DOI:10.19850/j.cnki.2096-4706.2022.06.025
中图分类号：TP181 文献标识码：A 文章编号：2096-4706（2022）06-0100-04
Deep Subspace Clustering Method Based on the Multi-level Structure
YU Wanrong
(School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214122, China)
Abstract: A new deep subspace clustering method that uses a convolutional autoencoder to transform an input image into a representation that lies on a linear subspace is proposed. The feature learning process is facilitated by combining low-order and highorder information extracted by the autoencoders, and multiple sets of self-representations and information representations ar e generated at different levels of the encoder. The obtained multi-level information is fused to obtain a unified coefficient matrix and use it for subsequent clustering. The effectiveness of the above innovations is verified through multiple experiments on three classic datasets, including Coil20, ORL and Extended Yale B. And the clustering accuracies reach 95.38%, 87.25% and 97.58% respectively. Compared with other mainstream methods, this method can effectively improve the clustering accuracy and it has strong robustness.
Keywords: subspace clustering; multi-level structure; autoencoder
参考文献：
[1] 张红颖，贺鹏艺，王汇三 . 基于改进 SiamFC 的实时目标跟踪算法 [J]. 激光与光电子学进展，2021，58（6）：308-316.
[2] 黄志标，姚宇 . 基于像素聚类的超声图像分割 [J]. 计算机应用，2017，37（2）：569-573.
[3] 邹武合，张雷，戴宁.基于关键姿势的单次学习动作识别[J].半导体光电，2015，36（6）：999-1005.
[4] 张爱华，雷小亚，陈晓雷，等 . 基于细胞神经网络的快速手语视频分割方法 [J]. 计算机应用，2013，33（2）：503-506.
[5] JI P，ZHANG T，LI H D，et al.Deep Subspace Clustering Networks [C]//31st Conference on Neural Information Processing Systems （NIPS 2017）.Long Beach：CURRAN ASSOCIATES.INC，2017：24-33.
[6] ELHAMIFAR E，VIDAL R. Sparse Subspace Clustering：Algorithm，Theory，and Applications [J].IEEE Transactions on Pattern Analysis and Machine Intelligence，2013，35（11）：2765-2781.
[7] LIU G，LIN Z，YAN S，et al. Robust Recovery of Subspace Structures by Low-Rank Representation [J].IEEE Transactions on Pattern Analysis & Machine Intelligence，2013，35（1）：171-184.
[8] LU C Y，MIN H，ZHAO Z Q，et al. Robust and Efficient Subspace Segmentation via Least Squares Regression [C]// Proceedings of the European Conference on Computer Vision. Berlin，Germany：Springer，2012，7：347-360.
[9] XIE J，GIRSHICK R，FARHADI A. Unsupervised Deep Embedding for Clustering Analysis [C]//Proceedings of the 36th International Conference on Machine Learning. Long Beach：JMLR，2016：478-487.
[10] REN Y Z，WANG N，Li M X，XU Z L. Deep Densitybased Image Clustering [J].Knowledge-Based Systems，2020（197）：105841.
[11] PATEL V M，VIDAL R. Kernel Sparse Subspace Clustering [C]//Proceedings of the IEEE International Conference on Image Processing. Paris France：IEEE，2014：2849-2853.
[12] JI P，SALZMANN M，LI H D. Efficient Dense Subspace Clustering [C]//Proceedings of the IEEE Winter Conference on Applications of Computer Vision. Steamboat Springs：IEEE，2014：461-468.
作者简介：郁万蓉（1997.08—），女，汉族，安徽蚌埠人，硕士研究生在读，主要研究方向：子空间聚类、模式识别。