Attribute-based Classification For Zero-shot Visual Object Categorization By Low-rank Representation

Compared with traditional image recognition,zero-shot learning uses semantic knowledge migration of known class samples to predict the label of unknown category samples.Attribute-based zero-shot learning is to realize the knowledge transfer between the known category sample and the Unknown class sample by sing the attribute as an intermediate representation of the image sample.The existing zero-shot learning algorithms based on attributes have the problem of semantic migration and the noise of attribute and so on.In order to improve these problems effectively,the main research contents are as follows:Firstly,based on the attribute sparse representation of the zero-shot learning algorithm(RKT),a zero-shot learning algorithm(ZSL)based on the attribute which was low rank representation is proposed.The sparse representation method is replaced by the method of low rank representation on the basis of RKT,because the attribute of the class is defined artificially and has a certain noise.In order to eliminate these noises,the method of low rank representation can effectively achieve this effect,thus improving the recognition rate of zero-shot learning.Secondly,we propose a zero-shot learning algorithm(TZSL)based on full training of attribute low rank representation.The algorithm firstly uses the algorithm of low rank representation to train the coefficients that the Unknown class attribute is linearly represented by the known class attribute.And then migrates the obtained coefficient into the feature space to make the sample of the unknown category.Finally,all the samples of the feature space Trains attribute predictors with all attributes in the attribute space.The algorithm effectively improves the recognition rate of zero-sample images.Finally,a zero-shot learning algorithm(CTZSL)based on attribute-extracted low rank representation is proposed,because some attributes are unreliable in the attribute space of zero-shot learning,or some attribute relationships are very close and cannot achieve the due effect.The method is to extract the attributes in the attribute space by removing the unreliable and useless attributes,and then use the low rank to represent the coefficients of the linear representation of the classes of unknown class attributes,and then use the obtained coefficients to dummy out the unknown category and train the property predictor.This method can effectively improve the recognition rate of zero-shot learning.In order to verify the effectiveness of the methods mentioned above,the comparison between the proposed method and the comparison method is carried out in the Animal Attributes Database(AWA),the California Bird Database(CUB)and the Yahoo Database(aP/aY).The experimental results also prove that the proposed method is superior to other zero-shot learning algorithms.