Research On Zero-shot Learning Methods Based On Attribute

Zero-shot learning aims to accurately recognize objects whose instances not have been seen during training.It makes the computer have the ability of knowledge transfer,which is very important for exploring and realizing real artificial intelligence.In order to transfer knowledge from training class to testing class,zero-shot learning needs the intermediate representation of attribute.In this paper,four zero-shot learning methods based on attribute are proposed through the research and exploration of such issues as attribute relation learning,attribute feature learning,domain shift and space shift.The main work includes:1.Aiming at the problem that each attribute is individually trained and the relation between attributes is neglected in the classical zero-shot learning methods represented by direct attribute prediction model,a zero-shot learning method based on the attribute relation graph regularization is proposed.Firstly,the correlations between attributes are calculated by correlation coefficient based on the class-attribute matrix and the attribute relation graph is constructed.Then,the attribute learning is regard as a multi-task learning and the attribute relation graph regularization is integrated into multi-task attribute learning to learn multiple attribute classifiers.Finally,zero-shot classification of test samples is performed according to the learned attribute classifiers.2.Aiming at the shortcomings of multi-task attribute learning in attribute relations and shared features by attributes,a zero-shot learning method based on mixed attribute relations and attribute specific features is proposed.Firstly,we assume that each attribute has its specific features and the number of shared features between attributes is related to the strength of the second-order attribute relation,attribute features learning based on the second-order attribute relation is studied.Then,the high-order attribute relation is constructed by sparse representation and is unified into the multi-task attribute learning based on the second-order attribute relation.The attribute classifiers based on the second-order attribute relation are modified to get the final attribute classifiers.Finally,zero-shot classification is performed according to the final attribute classifiers.3.In order to solve the problem of domain shift in zero-shot learning from feature space to attribute space,a zero-shot learning method based on feature prototype is proposed.Firstly,feature prototype of training class is studied according to the features and the class labels.Secondly,the linear and nonlinear modeles are built between the class feature prototype and the class attribute.Then,the feature prototype of test class is synthesized according to the modeles and the test class attribute vector.Finally,zero-shot classification is performed by comparing the distance between the test class sample features and the class feature prototype.4.In order to solve the problemes of domain shift and space shift in generative zero-shot learning,zero-shot learning method based on coupled autoencoder and gaussian mixture model is proposed.Firstly,we assume that both the training class and the test class obey the gaussian distribution in the feature space.Secondly,the class conditional distribution of the training and test classes are jointly modeled by coupling autoencoder to reduce the domain shift.Then,the generated test class conditional distribution is optimized by the gaussian mixture model to rectify the space shift.Finally,the rectified class conditional distribution is classified by the maximum posterior probability.There are 53 figures,16 tables,and 148 references in this dissertation.