Image Relative Attribute Learning And Application

With the rapid development of the electronic information technology,the popularity of smart devices and the acceleration of network allow us to generate or acquire electronic document,image and video anytime and anywhere.With the electronic data on the Internet more and more,it's becoming a big challenge for people to understand and use these datas.Because expression information by image is much more clearly than document,and images save more storage space than videos,there are lots of images on the internet than others.Computer vision becomes a hot topic in current researches,and effective understanding images have been a foundation of lots of applications.There is a semantic gap between image features and semantic description.In order to eliminate this problem,some researches put forward the concept of the visual attribute.Attribute,as a kind of middle semantics,communicate the low-level features and high-level semantic and be used in image classification,image retrieval,target recognition,etc.The research on attributes is always divided into two branches that are binary attributes and relative attributes.Binary attributes indicates whether the image has a certain attribute,but sometime the judgment is not accurate.Relative attributes indicates the strength of attribute between images,it's more intuitive and can provide much more information.So it has been used widely.In this article,we mainly concern about two aspects: how to improve the accuracy of relative attribute model learning and how to use relative attribute in zero-shot image classification.Firstly,visual features extracted from images are been used in relative attribute model learning.This paper analyses the problem that features and attributes do not match.And put forward a new method can choose the matching features for attribute learning.Through made the features becoming sparse,improving features generalization ability for attributes.Based on this idea,alter the relative attribute learning equation and optimization.This paper tests the model on five databases on different aspects.The result shows that compared with the original feature,the selected feature can improve the model accuracy.Secondly,in the process of supervised learning,the labelled samples are needed to train the model.But in some cases,the labelled samples are not enough to cover all the classes.And zero-shot image classification is able to solve this problem,so it's becoming a hot topic in the field of transfer learning.To realize the transfer of attributes from seen to unseen classes in zero-shot image classification,the classification model needs to build.First,transform the image from low-level features to relative attributes features,then make a Gaussian model to describe the images in the same class.Using the relationship between seen class and unseen class,get the unseen class Gaussian model.Finally,the test is carried out on two databases,and the experiment result demonstrates that the model can improve the accuracy of zero-shot image classification.