摘要：

现有的图像集分类算法在进行图像集表示时往往做出多种假设,无法有效描述图像集的特点,且难以利用图像集中区分性信息进行分类。为此,借鉴深度学习的思想,提出一种改进的图像集分类算法。将每个图像集模拟为一个流形并作为多层深度神经网络的输入,通过激励函 数使得各个流形非线性地映射到另一个特征空间。在网络的最顶层,采用反向传播和最大流形边缘准则训练和优化流形的参数。在测试阶段,使用训练得到的深度网络,计算测试图像集和所有训练类别之间的相似性,并利用最短距离进行分类。实验结果表明,与判别典型相关 分析算法、流形到流形距离等算法相比,所提算法的分类精度更优、运行时间更短。

关键词: 图像集分类, 神经网络, 激励函数, 流形边缘, 最短距离

Abstract:

The existing image set classification methods often make many assumptions for representing images,cannot effectively describe the characteristics of the image set,and it is difficult to classification using the discriminative information in the image set.To solve this problem,an improved image set classification algorithm is proposed by using the idea of deep learning.Each image set is modeled as a manifold and is used as input to the multi-layered neural network,and then the manifold is mapped nonlinearly to another feature space via the excitation function.At the top of the network,the parameters of the manifold are trained and optimized by back propagation and maximum manifold margin criterion.In the testing phase,the training network is used to calculate the similarity between the test image set and all the training classes,and the shortest distance is used to classify the test image.Experimental results show that compared with Discriminant Canonical Correlation(DCC),Manifold-Manifold Distance(MMD) and other algorithms,the proposed algorithm has better classification accuracy and shorter running time.

Key words: image set classification, neural network, excitation function, manifold margin, shortest distance

中图分类号: 

• TP393