Zero-shot Image Classification Based On Attribute Extension

As the labeled training samples are inadequate to cover all the object classes,zeroshot image classification is to distinguish the object classes which have not any labeled samples in the training stage.It is impossible for traditional classifiers to recognize new object classes correctly.For zero-shot image classification based on attribute learning,attributes are used to describe the semantic information and the visual information of the object.Furthermore,as the sharable intermediate layer,attributes can transfer knowledge from the ‘seen' classes to the ‘unseen' classes.However,the classification performance of zero-shot image classification is seriously influenced by the description and coverage of the attributes.From this viewpoint,the contents of the thesis are arranged as follows:Firstly,a novel adaptive zero-shot image classification based on extended relevant non-semantic attributes model is proposed.First,elastic net regularization and dictionary learning with two-stage optimization are used to reconstruct the features of the objects.The relationship between semantic attributes and non-semantic attributes is mined to obtain the relevant non-semantic attributes.The relevant non-semantic attributes of objects are used to improve the ability of the spatial representation.Second,the combination of semantic attributes and non-semantic attributes constructs an enhanced attribute representation space and extends the coverage of attributes.Then,domain adaptation and dictionary learning are combined to minimize the difference between the training domain and the testing domain on the sharable attribute space.Finally,the class label is predicted by the projected relations between the attributes and the different classes.Secondly,a novel zero-shot image classification based on multi-task extended attribute-grouping learning model is proposed.First,the class labels are regarded as the extended attributes to improve the coverage of the attribute space.Second,for training sample,combing the structured sparse method,the multi-task learning model is built on the attributes grouping and the classes grouping.The relationship between classes and between attributes is discovered.Then,Jaccard similarity coefficient is applied to discover the correlations between the classes(training classes and testing classes).And the relational matrix is obtained.Finally,the zero-shot image classification is achieved by constructing the full-connected model of the class-attribute-feature.The experiments are conducted on OSR,Shoes and AWA datasets.The results show that the proposed methods can yield better zero-shot image classification performancethan other comparative methods.