Research On Zero-Shot Learning Algorithm Based On Deep Learning

Image classification methods based on deep learning have achieved remarkable achievements in multiple intelligent recognition fields.However,with the use of deep learning,its limitations gradually appear.It is necessary to ensure a large amount of labeled training data for each category.However,many practical situations are difficult to meet this.So the emergence of zero-shot learning solves this problem.Zero-shot learning based on deep learning can not only solve the problem of lack of labeled data,but also have great significance for exploring true artificial intelligence.This article does a lot of research on zero-shot learning.The main contributions are as follow:(1)Aiming at the problem of insufficient binding and robustness of deep regression models with semantic features embedded in image space,a zero-shot learning algorithm based on semantic alignment and reconstruction is proposed.Basic model is a deep regression model with semantic features embedded in the image space.In view of the problem that the embedded semantic features tend to tend to the origin of the space and become the pivot point,semantic alignment constraint is proposed to bring the embedded semantic features and the center point of the image features closer,thereby reducing the offset of the semantic features and further alleviating the impact of hubness on the classification accuracy.In addition,for the problem of insufficient robustness,semantic reconstruction constraint is proposed,so the model can retain complete and correct semantic feature information,thereby improving the generalization of the model and further alleviating the impact of domain drift on the classification accuracy.Through experiments on the Aw A(Animals with Attributes)and CUB(Caltech UCSD Birds-200-2011)data sets,we can see that compared with other algorithms,the algorithm in this paper has obtained a high classification accuracy rate on Aw A,and it is also obtained a relatively good classification effect on CUB.In addition,through experiments on the two constraints mentioned,it is found that the model with added constraints has a higher classification accuracy than the basic model,and the effectiveness of the constraints in improving the model's zero-shot classification accuracy is verified.(2)Aiming at the problem of the comparison metric module of the relational network ignoring the impact of the quality of embedded semantic features on the accuracy of relationship scores,distance matching constraint is proposed to ensure the accurate distribution of embedded semantic features,thereby improving the accuracy of relationship scores.At the same time,for the single auxiliary information of the relationship network,there is a limitation to the description of categories.Adding other auxiliary information and using neural networks to merge multiple semantic information to enhance the richness of semantic knowledge.To this end,a zero-shot learning algorithm based on multi-semantic improved relational network is proposed.Aiming at the problem that the CUB features extracted by the previous convolutional networks ignore the phenomenon of small inter-class differences and large intra-class differences of fine-grained images,bilinear convolutional network is introduced to extract fine-grain image features,thus improving the judgment ability between different categories.To this end,a zero-shot learning algorithm based on bilinear convolutional networks and multisemantic improved relational networks is proposed.The two proposed algorithms are experimented on Aw A and CUB respectively.Compared with other zero-shot algorithms,two algorithms can improve the classification accuracy of Aw A and CUB respectively.Through experiments on the proposed distance matching constraint and multi-semantic information,we verified their effectiveness in improving classification accuracy.In addition,the visual analysis of the bilinear features of CUB and other convolutional network features proves that fine-grained image features extracted by the bilinear convolutional network have better intra-class convergence and distinguishability between classes.