基于自组织增量-图卷积神经网络的金相图半监督学习

采用深度学习对钢铁材料显微组织图像分类，需要大量带标注信息的训练集。针对训练集人工标注效率低下问题，该文提出一种新的融合自组织增量神经网络和图卷积神经网络的半监督学习方法。首先，采用迁移学习获取图像数据样本的特征向量集合；其次，通过引入连接权重策略的自组织增量神经网络(WSOINN)对特征数据进行学习，获得其拓扑图结构，并引入胜利次数进行少量人工节点标注；然后，搭建图卷积网络(GCN)挖掘图中节点的潜在联系，利用Dropout手段提高网络的泛化能力，对剩余节点进行自动标注进而获得所有金相图的分类结果。针对从某国家重点实验室收集到的金相图数据，比较了在不同人工标注比例下的自动分类精度，结果表明:在图片标注量仅为传统模型12%时，新模型的分类准确度可达到91%。

Semi Supervised Learning of Metallographic Data Based on Self-organizing Incremental and Graph Convolution Neural Network

It needs a large number of training sets with annotation information to classify microstructure images of steel materials by deep learning. To solve the problem of low efficiency of manual image annotation, a new semi-supervised learning method combining self-organizing incremental neural network and graph convolutional neural network is proposed. Firstly, it uses transfer learning to obtain the feature vector set of images. Secondly, it obtains the topology structure by adopting the Weighted Self-Organizing Incremental Neural Network(WSOINN) based on connection weight strategy to learn feature data, and manually annotates a small number of nodes which are selected by the number of victories of node. Then, a Graph Convolution Network (GCN) is built to mine the potential connections of nodes in the graph, dropout is used to improve the generalization ability of the network, and the remaining nodes are automatically annotated to obtain the classification results of the metallograph. Experiment on the metallographic data collected from a state key laboratory, the accuracy of automatic classification under different manual annotation ratio is compared. The results show when the image annotation amount is only 12% of the traditional model, and the classification accuracy of the proposed model can reach up to 91%.