Zero-shot Image Classification Based On Semantic Attribute

In computer vision and pattern recognition,there exists a problem that the labeled training samples can not cover all classes,which is named zero-shot problem.Because of the significant difference in distributions between the training class and the testing class,the traditional classifier can not be directly used for zero-shot learning.Based on the problem,in recent years,attribute learning has been widely concerned.Different from the low level feature of the image,the attribute is the high level description of the image content,which can be understood by machines and persons simultaneously.Therefore,the attribute can be used as a shared intermediate representation between the training class and the testing class,which enables the knowledge to transfer from the training classe to the testing classe.At present,there are two shortcomings in attribute learning: 1)the training samples have the same influence on the classification,the trained model can not satisfy the testing class which results in the lower classification accuracy;2)the different attributes have the same influence on the classification which affect the attribute prediction accuracy of the model.Based on the above defects,the main contents of this thesis are as follows:(1)In the traditional attribute learning model,it is assumed that the contribution of each sample to the training model is same.But,the mutual exclusion or distributional difference between training and testing classes result in that the trained model can not satisfy the testing samples.Therefore,an indirect attribute prediction(IAP-SaW)model based on sample weights is proposed.First,the mean value of the low-level feature of the testing samples is obtained as the reference sample of the testing class.Then,the cosine similarity and European distance between the training samples and the reference sample are calculated separately,the ratio of which is taken as the weight of the sample.Finally,the sample weight is introduced into training the indirect attribute prediction(IAP)model and the IAP-SaW model is obtained for zero-shot image classification.(2)In IAP model,it is assumed that each attribute of the sample is same to the classification decision,which means that each attribute's weight equals to one.However,when training the attribute model,the attribute prediction model based on this assumption has lower image recognition rate because of the inaccurate prediction attribute.Therefore,an indirect attribute prediction(IAP-AAW)model based on adaptive of attribute weight is proposed.First,the relationship between the category and its attributes is used to calculate the initial weight of the attribute.Then,the weight is combined with the trained attribute model to classify the training samples,and the error recognition rate is obtained which is used to adjust the attribute weight.Finally,the weight is introduced into IAP model to achieve the label prediction of the testing sample.Experiments are performed on the multiple datasets.And compared with the traditional attribute prediction model,the proposed methods obtain higher accuracy for zero-shot image classification.