Zero-shot Image Recognition Based On Auto-encoder

In zero-shot image recognition,there are two sets of instances belonging to disjoint classes.One set contains labeled source instances of seen classes and the other set contains unlabeled target instances to be classied in new classes that have never been seen before.zero-shot image recognition transfers the knowledge with semantic attributes from the source set to the target set,thereby identifying the labels of the samples in the target set.The traditional zero-shot image recognition algorithm directly applies the recognition trained in the source set to the target set.Such methods cause domain shift problem;in addition,the traditional zero-shot recognition algorithm considers that the semantic attributes of all samples from the same class can be representated by the prototype attributes of the class.However,the diversity of the attribute information of samples will loss due to that the difference between semantic attributes and prototype attributes is unignorable.In view of the above problems,the main work is as follows:Firstly,to address the domain shift issue,we propose a novel zero-shot image recognition algorithm based on attribute-constrained auto-encoder(ACAR).(1)In source set,we learn a projection from image feature space to semantic attribute space using auto-encoder.Since the projection of target set is unknown,the value of the source set projection is considered as the initial value of the target set projection.(2)Using auto-encoder in the target set,the auto-encoders of source set and target set can be associated by adding regularization about the two domain projections.At the same time,the attribute constraint item of the unseen class samples is added to the auto-encoder of the target set,and the attributes of the unseen class samples are obtained by an iterative algorithm.(3)The cosine similarity is used to predict the labels of the unseen class samples.Secondly,on the basis of ACAR algorithm,we propose a zero-shot image recognition algorithm by adopting the principle of coupled auto-encoder(CALC).(1)Using auto-encoder in the source set,we add an additional constraint,which consists of two parties,one is the attributes matrix constructed from the source set projection matrix and the sample features,and the other is the prototype attributes matrix of the class.(2)Using auto-encoder in the target set,this step is the same as ACAR,and the attributes of the unseen class samples are obtained by the iterative algorithm.(3)The cosine similarity is used to predict the label of the unseen class sample.Thirdly,to alleviate the negative effects of dipcting each sample with the prototype attribute of the corresponding class,we propose a zero-shot image recognition algorithm based on the attribute auto-encoder(TALC).(1)The attributes of seen class samples are learned in the source set by using auto-encoder.(2)We use the learned attributes to replace the original attributes of the seen class samples.We use the CALC algorithm to learn the attributes of the unseen class samples.(3)The cosine similarity is used to predict the label of the unseen class sample.In order to verify the effectiveness of the proposed algorithms,we perform experiments on the Animal with Attribute(AwA),Caltech-UCSD Birds 2011(CUB)and aPascal & aYahoo(aPY).Experiments show that the proposed algorithm is superior to some competitive zero-shot recognition algorithms.