Research On Zero-shot Recognition Algorithm Based On Domain Adaptation

Zero-shot Recognition can be regarded as a special kind of transfer learning.The particularity is that the training sample set and the test sample set have different labels and the samples in the test sample set are unlabeled.The traditional Zero-shot Recognition algorithm learns from the training sample set to identify the model.Then apply it to the test sample set without any adaptability.The above Zero-shot Recognition algorithm generally has the following deficiencies.First,in the learning phase,the discrepancy of sample distribution between the training sample set and the test sample set is not considered,which leads to the occurrence of projection domain transfer.Second,in the learning phase,the attribute classifier learned is not suitable for test sample set because it does not take into account that each prototype attribute does not represent the attribute representation of all samples in the class.Thirdly,in the learning phase,since the similarity information retention between similar sample features is not considered,the attributes of the same type of samples are separated from each other in the attribute space.Therefore,it is of great significance to study how to alleviate the impact of the above problems on Zero-shot Recognition algorithm.The main work is as follows:Firstly,in response to the above first and second questions,we propose an adaptive sample attribute(ALIA)algorithm.Above all,we learn the attribute classifier from the feature space to the attribute space on the training sample and the test sample set.Then,using the class prototype of the training sample set to adjust the learning attribute classifier to make it the best.Finally,we use this attribute classifier to predict the attribute representation of the test sample for recognition.In addition,this algorithm also learns the sample attributes in the training sample set to alleviate the influence of the above second problem on the final recognition performance.Secondly,on the basis of the ALIA algorithm,we propose a discriminative sample attribute(LDIA)algorithm.Above all,we use the indirect attribute prediction algorithm to get the pseudolabel information of the test sample as the weight matrix.Then,the iterative process updates the weight matrix through learning the class prototype attribute until it converges to obtain the optimal attribute classifier.Finally,we use the learned attribute classifier to get the attribute of the test sample to identify.Thirdly,in response to the third question,we propose a discriminative and predictable sample attribute(LDPIA)algorithm.Above all,construct the adjacency matrix by using class labels of the training sample set.Then,the congeneric sample similarity constraint regularization term is constructed by adjacency matrix.Finally,we use the learned attribute classifier to obtain the attribute representation of the test sample and then identify it in the attribute space.We evaluated the effectiveness of the proposed algorithms by performing experimental comparison on the AwA,Caltech-UCSD Birds 2011 and aPascal-aYahoo datasets.Experiments show that the proposed algorithm is superior to some competitive Zero-shot Recognition algorithms.