结构化深度判别嵌入编码网络图像聚类  

Structured Deep Discriminant Embedded Coding Network for Image Clustering

作　　者：付兴武[1] 吕明明 刘万军[1] 魏宪 Fu Xingwu;LüMingming;Liu Wanjun;Wei Xian(College of Software Liaoning Technical University,Huludao Liaoning 125105,China;Quanzhou Institute of Equipment Manufacturing Haixi Institutes,Chinese Academy of Sciences,Quanzhou,Fujian 362200,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125105 [2]中国科学院海西研究院泉州装备制造研究所,福建泉州362200

出　　处：《激光与光电子学进展》2021年第6期233-243,共11页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61806186)。

摘　　要：大部分现有深度聚类方法都试图最小化重构损失,然而深层特征的判别能力与重构损失并没有必然联系,并且这些深度聚类方法通常只关注从样本自身提取的有用特征,很少考虑样本背后的结构信息。为解决这些问题,提出一种新的结构化深度判别嵌入编码网络聚类(SDDECC)算法,用于无监督图像聚类。首先在多层卷积自编码器网络中引入最大化互信息与最小化先验分布约束,然后使用传递算子将深度判别嵌入编码网络(DDECN)模块学习到的特征表示融入到图卷积神经网络(GCN)模块中,最后利用Kullback-Leibler(K-L)散度联合双网络结构产生的潜在特征分布端到端地完成聚类训练。实验结果表明,SDDECC算法能够有效提取更多有鉴别性的深层特征,并且由于在GCN中融合了样本的属性信息和结构信息,最终该模型取得了良好的聚类效果。Most existing deep clustering methods are employed to minimize the reconstruction loss.However,the identification ability of potential representation is not necessarily related to the reconstruction loss.Moreover,these deep clustering methods focus only on extracting useful features from the sample itself and seldom consider the structure information behind the sample.To resolve these problems,a new structured deep discriminant embedded coding network clustering(SDDECC)algorithm is proposed for unsupervised image clustering.First,the maximum mutual information and minimum prior distribution constraints are embedded in a multilayer convolutional autoencoder network.Then,the feature representation learned by the deep discriminant error correction network(DDECN)module is integrated into a graph convolutional neural network(GCN)module by the transfer operator.Finally,Kullback-Leibler(K-L)divergence is used in combination with the potential feature distribution generated by the dual network structure and is trained end-to-end to guide the clustering learning.The experimental results show that SDDECC algorithm can effectively extract more discriminative deep features than those obtained using traditional methods.Moreover,because the attribute information of the sample itself and the structural information between the samples are integrated in the GCN,the model shows good clustering.

关 键 词：图像处理 深度聚类 图卷积神经网络 无监督学习 三元组互信息 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]