Research On Zero-Shot Learning Combining Autoencoder And Knowledge Graph

In recent years,deep learning has achieved good results in many fields,but its biggest drawback is that it requires a large amount of labeled training data during the learning process,which will consume a lot of time and energy.Zero-shot learning can better solve this problem.It uses the marked visible class training set and knowledge about the invisible class and the semantic related knowledge of the visible class to enable the model to identify the class samples that did not appear in the training process.There are visual features and semantic features in zero-shot learning.Whether these two features can better represent the corresponding category has a great influence on the final classification result.Therefore,the optimization of the feature itself is very important,especially the semantic feature,because the semantic feature of the category is mostly artificially defined,and there is inevitably the problem of inaccurate image description.In addition,zero-shot learning also needs to map semantic features and visual features to the same dimensional space.This is the key in zero-shot learning.Different mapping methods have a great influence on the final result.Based on the research and analysis of the existing work,this paper proposes the related dual autoencoder method and the zero-shot learning method based on knowledge graph.The main research contents are as follows:(1)Aiming at the problem that semantic features are not accurate enough for image description,this paper proposes a model architecture of related dual autoencoders.This paper establishes autoencoders for visual features and semantic features respectively.These two autoencoders are related.The results produced by the visual feature autoencoders will affect the encoding and decoding process of the following semantic feature autoencoders.In this way,the visual information of the picture can be included in the semantic feature,and the semantic feature can describe the picture in a more complete and fine-grained manner,thereby playing a role in improving the recognition classification accuracy.(2)The connection relationship between categories can help to learn some hidden semantic features,so as to better map with visual features,so this paper uses the Word Net knowledge graph to construct the relationship graph between categories and introduce additional prior knowledge.Then use the graph network algorithm Graphsage to learn and train the relationship graph,and constantly absorb and learn the information of adjacent nodes,so that the nodes in the relationship graph have a more complete semantic description.Because of the introduction of graph structure,this paper uses graph convolutional neural networks to map semantic features to the same dimensional space as visual features.In addition,this paper also uses the transductive learning to introduce the pictures from the test set into the training process,so that the learned knowledge can be better transferred from the training set category to the test set category and overfitting can be reduced.In this paper,these two models have been extensively tested on the Aw A,CUB,and Image Net datasets,and have achieved better results than some previous models.The second model method has better performance.