Research On Zero-Shot Classification Based On Common Space Embedding

In recent years,with the development of deep learning technology,methods based on supervised learning in the field of image recognition have achieved significant performance improvements,and image recognition algorithms have become more and more powerful.In order to obtain the excellent accuracy,a large number of labeled training samples have to be provided for each class in the dataset.However,some data labels are sometimes difficult to obtain,or a large amount of data needs to be manually labeled.In addition,the number of object types is still showing an increasing trend,which requires the recognition system to increase and rebuild new data continuously.The zero-shot classification algorithm has been widely concerned in recent years,in which the labeling of samples of a new category is unnecessary and the cost of annotations can be reduced in applications.Therefore,the classification method based on zero-shot learning algorithm came into being.The purpose of zero-shot learning is to solve the learning task that lacks labeled data.Based on the analysis of the current mainstream zero-shot learning algorithms,three improved zero-shot classification models are proposed in this paper:1.A zero-shot learning algorithm model based on deep common space is proposed.The model uses deep learning technology to map image feature vectors and semantic feature vectors to the common space by adjusting the network structure.The distance between the two vectors is measured by the distance metric learning(DML)method,and the classification result is given according to the distance.It establishes an end-to-end deep image feature extraction and common space embedding model.Therefore,the algorithm can train the parameters of the image mode and the semantic modal simultaneously,which improves the performance of accuracy.2.Based on the previous model,this paper presents a zero-shot classification method based on word vectors enhancement and distance metric learning.The analysis dictionary learning(ADL)method is implemented in sparse representation of word vectors to alleviate redundant information.The objective function of the ADL model is improved,and a LC-ADL model combining with a synthetic linear classifier is proposed.It further reduces noise and errors from word vectors.In the distance measurement module,the large margin nearest neighbor(LMNN)algorithm of DML method is introduced.Reconstructing the loss function can effectively reduce the error rate and the computational complexity.It has better applicability.3.A zero-shot classification method based on multi-semantic features and common space embedding is proposed.Both attributes and word vectors are utilized to further improve the classification accuracy.The output features are considered as structured objects in this algorithm.With the introducing of nonlinear mapping framework by the structured embedding,test images can match semantic features better according to a compatibility function.Therefore,the algorithm has better generalization ability.Comparison experiments are performed on the Aw A2,CUB and A-Pascal/A-Yahoo datasets to verify the feasibility and effectiveness of the proposed algorithms.