Zero-shot Image Classification Based On Generative Model

Zero-shot learning(zero-shot image classification)mainly studies the use of additional information such as attributes or word vectors as a semantic space,and how to correctly classify test samples without the intersection of training categories and test samples.And how to reduce the problem of domain shift caused by the distribution difference between the training and the test samples.This dissertation mainly studies the classification of zero-shot based on the perspective of generative models.The main contents of this dissertation are as follows:Firstly,aiming at the problem of domain offset when semantic attributes are mapped to image features in zero-sample classification,a generative zero-shot model combining autoencoders and class feature prototypes is proposed.This model first builds class feature prototypes of image features.As a marker for each category,then uses the kernel function to construct the attribute vector kernel matrix,and then uses the autoencoder to model the class feature prototype and attribute kernel matrix of the training category.The auto-encoder reconstructs the image features and encodes them first.After decoding,two mappings can reduce the information loss between features and semantic attributes.Then apply the attribute kernel matrix of the test class to the autoencoder model to generate the class feature prototype of the test class,and then uses the K nearest neighbor samples near the class feature prototype of the test class to modify the generated test feature feature prototype,and finally in the image feature space uses the nearest neighbor classifier for the zero-shot classification of the modified feature prototype.Secondly,for most zero-shot learning based on generative models,no consideration is given to the existence of noise in the generated test samples and the inconsistency with the feature distribution of the real test samples.A zero-shot learning model based on conditional variational encoder and domain adaptation is proposed.The model first fuses the image features and semantic attributes of the training class as the input of the model,uses a variational autoencoder network to mine the relationship between the semantic attributes and the image features,and reconstructs the input vector.After the model training is completed,then The image features of the test class are initialized to a Gaussian distribution,and samples of the test class are generated using a variational autoencoder.In order to adapt the distribution of training classes and test classes to make them consistent,while reducing the characteristic noise of the generated samples,a domain adaptive method was introduced to filter out features similar to the test classes,and then the noise samples were removed.Finally,the generated test classes were used to train the SVM classifier for zero-shot classification.In this dissertation,simulation experiments are performed on the AWA1,AWA2,and CUB datasets,and theoretical analysis and exploration of the experimental results are performed,which confirms that the model in this dissertation has good classification performance compared to other models.The dissertation contains 36 pictures,8 tables,and 100 references.