| In recent years,with the development of deep learning technology,methods based on supervised learning have seen huge performance improvements.Especially in the image recognition direction,its accuracy has exceeded the human recognition ability.In order to obtain this excellent accuracy,a large number of labeled training samples need to be provided for each class in the dataset.However,because the classifier learned by the supervised method cannot be well transferred to the image set of other classes,every time a new scene is encountered,a labeled dataset that requires a lot of labor needs to be created.The need for labeled datasets has greatly hindered the development of supervised learning.How to make the model applicable in the new environment to reduce the labor consumption in the production of image data sets has become an urgent problem to be solved.The purpose of zero-shot learning is to solve this kind of learning task lacking labeled data.Although current methods based on zero-shot image classification have made good progress,there are still many areas for improvement.For example,the model can not effectively use the abundance of priori knowledge,and the semantic gap caused by the mismatch between the semantic space and the visual space,and so on.Therefore,in order to solve these problems that exist in the zero-shot image classification task,methods from the network model and the classification characterization definition are proposed,mainly including the following work:(1)a new relationship-enhanced graph convolutional network model based on visual attributes and graph attention mechanism is proposed.The introduction of graph-based attention mechanism and relationship-enhanced knowledge graph strengthens the relationship between categories to better achieve semantic knowledge transfer between classes.(2)Aiming at the problem of semantic interval in zero-shot learning,an algorithm based on semantic information fusion is proposed,which effectively combines word embedding vectors and visual feature descriptions to reduce the impact of semantic interval.The method proposed in the thesis is applied to commonly used image classification datasets,and related experiments are conducted to verify the effectiveness of the proposed model. |
