融合关系特征的半监督图像分类方法研究

半监督深度学习模型具有泛化能力强，所需样本数较少等特点，经过10多年的发展，在理论和实际应用方面都取得了巨大的进步，然而建模样本内部“隐含”关系时模型缺乏解释性以及构造无监督正则化项难度较大等问题限制了半监督深度学习的进一步发展。针对上述问题，从丰富样本特征表示的角度出发，构造了一种新的半监督图像分类模型—融合关系特征的半监督分类模型（semi-supervised classification model fused with relational features，SCUTTLE），该模型在卷积神经网络模型(convolutional neural networks，CNN)基础上引入了图卷积神经网络(graph convolutional networks，GCN)，尝试通过GCN模型来提取CNN模型各层的低、高级特征间的关系，使得融合模型不仅具有特征提取能力，而且具有关系表示能力。通过对SCUTTLE模型泛化性能进行分析，进一步说明了该模型在解决半监督相关问题时的有效性。数值实验结果表明，三层CNN与一层GCN的融合模型在CIFAR10、CIFAR100、SVHN 3种数据集上与CNN监督学习模型的精度相比均可提升5%~6%的精度值，在最先进的ResNet、DenseNet、WRN（wide residual networks）与GCN的融合模型上同样证明了本文所提模型的有效性。

Semi-supervised image classification method fused with relational features

A semi-supervised deep learning model exhibits great generalization ability with minimal required samples and has made great progress in theory and practical application over the past ten years or so. However, the lack of the model’s interpretability when modeling the internal “implicit” relationship of samples and the difficulty in constructing unsupervised regularization items have limited the further development of semi-supervised deep learning. To solve these problems and enrich the sample feature representation, this study has developed a novel semi-supervised model for image classification—semi-supervised classification model integrating the relational features (SCUTTLE). The model introduces the graph convolutional networks (GCN) based on the convolutional neural networks (CNN) and extracts the relationships between the low- and high-level features of each layer of the CNN model via the GCN model, thus extracting features and expressing relationships. By analyzing the generalization performance of the SCUTTLE model, the paper further illustrates its effectiveness in solving semi-supervised related problems. The numerical results indicate that the classification accuracy of the hybrid model with three layers of CNN and one layer of GCN can be improved by 5%–6% compared to that of the CNN model on the CIFAR10, CIFAR100, and SVHN datasets. The effectiveness of the model proposed in this paper is also proved in the most advanced fusion models of ResNet, DenseNet, WRN (wide residual networks), and GCN.