Zero-shot Image Classification Based On Attribute Mining

Images are important carriers of information and data,which have widely infiltrated into all aspects of modern production and human activities.Using machine learning methods to classify and process massive images is becoming an urgent need for productivity improvement in various industry fields.Zero-shot image classification refers to predicting class labels of testing samples by attributes and other auxiliary information when the labeled classes in the training set cannot cover the whole samples in the testing set.At present,with the rapid growth of image classes and scenes,the continuous emergence of new classes and increasing requirements for classification fineness,zero-shot image classification exhibits a broad application prospect.Aiming at the limitations of current zero-shot image classification such as the incomplete construction of feature-attribute relation and the insufficient attribute description,four novel zero-shot image classification models based on attribute mining are proposed,where attribute-feature relation,attribute-attribute relation,attribute space structure and attribute-class relation are fully mined by employing machine learning methods such as the elastic net constraint,broad learning and attribute relation digraph.The main contributions of this dissertation are represented as follows.1.In view of the insufficient attribute and feature representation abilities in zero-shot image classification,a broad attribute prediction model with enhanced attribute and feature is proposed.Firstly,the elastic net constraint is utilized to learn binarized sparse enhanced attributes,which constitute hybrid attributes with manually labeled semantic attributes.Secondly,in order to expand existing image features,enhanced features are obtained through enhanced nodes of broad learning.Meanwhile,the model gains predicted results of all the attributes simultaneously by adopting the pseudo-inverse matrix of ridge regression in broad learning.Finally,the similarity between the predicted and hybrid attributes in testing classes is calculated by Manhattan distance,which is further used to implement image classification.2.Aiming at problems of the insufficient attribute representation and incomplete mapping between attributes and features,a novel zero-shot image classification method based on weighted reconstruction of hybrid attribute groups is proposed.Firstly,semantic attributes are grouped automatically by hierarchical clustering,and then are enhanced by group using the broad structure.The grouped semantic attributes,together with enhanced attribute groups,constitute the hybrid attribute groups.Secondly,to consider different contributions of hybrid attribute groups,the mapping between attribute space and feature space is constructed by weighted autoencoder.Meanwhile,the structured sparse L₂₁ norm is adopted in the objective function to remove attribute redundancy.Finally,the proposed method achieves zero-shot image classification by calculating the similarity between features of the testing sample and predicted class features in the feature space.3.In terms of current issues of incomplete mapping between attributes and features,as well as the insufficient consideration of the structure of attribute space,an attribute tri-factorization model with regularization of relation digraph is proposed.Firstly,the mapping between attribute space and feature space is achieved by matrix tri-factorization of attributes,and the obtained projection matrix is shared between training and testing stages.Secondly,the weight matrix is defined by weighting similarities between attributes,which is used to construct the attribute relation digraph.Finally,the similarity between the testing sample and each testing class is calculated in either the attribute space or the feature space to finish image classification.Aiming at the problem of projection domain shift,a transductive model is further proposed by simultaneously considering the relationship among testing classes and the distribution of testing samples.4.In order to make up for the lack of considering sample feature distribution in the relationship between attributes and features,a domain adaptation model for generative feature based on attribute kernel matrix is proposed.Our model takes into account the shortcomings of existing generative models that class attributes appear too similar and the distribution of generated samples is inconsistent with real samples in the testing set.Firstly,the kernel method is adopted to calculate the kernel function in the semantic space,which is further used to construct the attribute kernel matrix.Secondly,the semantic attribute-class relation matrix and the attribute kernel matrix are combined as the condition of the conditional Wasserstein generative adversarial network,based on which pseudo sample features are obtained.Thirdly,the joint distribution adaptation method is employed to reduce the difference of the marginal distribution and conditional distribution between labeled generated samples and unlabeled real samples.Finally,zero-shot image classification is achieved by supervised learning with generated samples in the testing set.Comparative experiments on public attribute datasets show that the proposed models effectively improve the accuracy of zero-shot image classification under different settings.The research results of this paper enrich the existing machine learning theories and methods,which therefore have important theoretical significance.At the same time,the proposed models have high practical value and can be easily applied to relevant industrial fields.A total of 58 figures are included in this dissertation,as well as 12 tables and 216 references.