MD-RBM神经网络模型及其在材料微结构中聚类研究

在传统的RBM神经网络的基础上提出一种新颖的MD RBM神经网络模型用于超高碳钢微结构高维图像数据的特征学习。该模型利用新的乘法距离(MD)取代欧式距离以计算高维图像数据之间的距离关系,有效缓解欧式距离在高维数据中的不稳定性问题。MD-RBM神经网络模型利用少量的成对约束监督信息引导其编码过程,使得一部分图像数据的隐藏层特征更加聚集在一起,而且同时使得一部分图像数据的隐藏层特征更加分散,由此得到高维图像数据的隐藏层特征表现出很好的聚类性能。实验选择两种经典聚类算法Affinity Propagation(AP)和Spectral Clustering(SC)作为对比,结果显示,基于MD-RBM模型的聚类识别算法比原始聚类算法、半监督算法以及基于RBM模型的聚类算法都表现出更优的聚类性能。

MD-RBM NEURAL NETWORK MODEL AND ITS CLUSTERING STUDY IN THE MATERIAL MICROSTRUCTURE

This paper proposed a novel MD-RBM neural network model based on traditional RBM for feature learning of high-dimensional ultra-high carbon steel(UHCS)microstructure images data.This model used a new multiplicative distance(MD)to replace the Euclidean distance to calculate the distance relationship between high-dimensional image data,and effectively relieved the instability of Euclidean distance in high-dimensional data.The MD-RBM neural network model used a small amount of pairwise constraint supervisory information to guide its coding process,which made the hidden layer features of some image data closer and the other part more remote.Therefore,the hidden layer features of high-dimensional image data showed better clustering performance.In the experiment,we chose two classical clustering algorithms Affinity Propagation(AP)and Spectral Clustering(SC)as comparison.The results show that the algorithm based on MD-RBM model has better clustering performance than the original clustering algorithm,semi-supervised algorithm and clustering algorithm based on RBM model.