Research On Zero-shot Classification And Retrieval Algorithm Based On Multidomain Features Fusion

Classification and retrieval of objects is one of the basic applications of machine learning.Many practical tasks can be considered as classification and retrieval tasks.Due to breakthroughs in deep learning techniques,the performance of large-scale labeled samples and large-capacity neural networks has greatly surpassed many traditional methods in many supervised learning tasks.However,this supervised classification method requires sufficient labeled samples,and the learned classifier cannot be extended to categories that are unseen in the training set,thus greatly limiting its application scenarios.In many applications,some categories may lack sufficient training samples or even no training samples.This problem is a zero-shot learning problem when a given sample instance is a category that is unseen in the training set.In recent years,the study of zero-shot learning has become a hot topic.This paper mainly studies the classification and retrieval of images in this zero-shot learning scenario.By studying and analyzing the zero-shot learning method based on semantic-visual relationship mapping model,this paper proposes a zero-shot learning method based on intermediate space classification,which effectively alleviates the projection domain bias encountered in zero-shot learning.Problems such as shifting and lack of discriminative space have achieved high performance in the classification and retrieval tasks of image classification.The main work and contributions of this paper are as follows:Firstly,this paper proposes a zero-shot learning method.In order to construct an intermediate space with sufficient discriminativeness,this paper uses a joint learning method.Information in different spaces can be recovered through the representation of the intermediate space and the corresponding base matrix.And the intermediate space is supercritically supervised by the corresponding label information.Secondly,considering that the information of different spaces may be complementary,this paper considers the information of multiple spaces to classify,and the similarity space and the fusion space have great differences.This paper combines the constructed similarity representation with the representation of the fusion space to further improve performance.Finally,this paper has carried out a more comprehensive experimental comparison,including zero-shot classification experiment based on attributes,zero-shot classification experiment based on word vector and zero-shot retrieval experiment,and the experimental results are analyzed.