Zero-shot Learning Image Target Recognition Based On Deep Learning

In recent years,with the rapid development of deep convolutional neural networks,image recognition methods based on supervised learning have made great progress.However,in order to achieve excellent classification accuracy,a large number of manual annotated data are needed,and the learnt classifier cannot effectively migrate to scenes with no annotated pictures.Therefore,how to reduce the manual work on data and how to make the model adapt to the emerging new samples quickly,become an urgent problem to be solved.Zero-shot learning is an effective solution to this problem,which aims to make the model have the ability to identify data categories that have never been seen before.At present,the resear-ch on zero-shot learning image recognition has made some progress in classification model and category definition,but there are still some shortcomings.On the one hand,the classification model itself is not expressive enough.On the other hand,the category definition does not make full use of the auxiliary information such as external description or knowledge.In order to solve these problems,this paper will make some improvements on classification model and category definition,mainly including the following innovations:(1)To solve the problem of insufficient constraints on mapping space in mapping-based classification model,a semantic mapping model SEMIIR which based simultaneously on inter-class and intra-class constraints is proposed in this paper,which optimizes the distance within and between classes at the same time.(2)In order to solve the ambiguity and non-visual problems in category definition based on word vectors,this paper proposes a semantic embedding algorithm VMSE based on visual modification,which modifies category vectors by introducing external descriptions.(3)Facing the problem of underutilization of prior knowledge,this paper proposes a residual graph convolutional network model ARGCN-DKGs with attention mechanism based on different types of knowledge graphs.By introducing residual mechanism and attention mechanism,and integrating different knowledge graphs,knowledge transfer between different categories can be realized.The three improved models proposed in this paper are applied to many standard datasets,and the experiment results verify the effectiveness and superiority of the proposed models in a variety of evaluation indicators.